diff --git a/.cargo/config.toml b/.cargo/config.toml
new file mode 100644
index 0000000..dd4dfb9
--- /dev/null
+++ b/.cargo/config.toml
@@ -0,0 +1,4 @@
+[build]
+# Limit parallel codegen units to reduce peak memory usage during compilation,
+# especially when building candle-flash-attn with CUDA.
+jobs = 2
diff --git a/.gitignore b/.gitignore
index 41db16e..4252e1a 100644
--- a/.gitignore
+++ b/.gitignore
@@ -8,3 +8,4 @@ __pycache__
 *.dsq
 baselines
 session
+build.log
diff --git a/Cargo.lock b/Cargo.lock
index 65cb2d8..9055ece 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -39,9 +39,9 @@ dependencies = [
 
 [[package]]
 name = "aho-corasick"
-version = "1.1.3"
+version = "1.1.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8e60d3430d3a69478ad0993f19238d2df97c507009a52b3c10addcd7f6bcb916"
+checksum = "ddd31a130427c27518df266943a5308ed92d4b226cc639f5a8f1002816174301"
 dependencies = [
  "memchr",
 ]
@@ -52,12 +52,6 @@ version = "0.2.21"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "683d7910e743518b0e34f1186f92494becacb047c7b6bf616c96772180fef923"
 
-[[package]]
-name = "android-tzdata"
-version = "0.1.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e999941b234f3131b00bc13c22d06e8c5ff726d1b6318ac7eb276997bbb4fef0"
-
 [[package]]
 name = "android_system_properties"
 version = "0.1.5"
@@ -69,9 +63,9 @@ dependencies = [
 
 [[package]]
 name = "anstream"
-version = "0.6.21"
+version = "1.0.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "43d5b281e737544384e969a5ccad3f1cdd24b48086a0fc1b2a5262a26b8f4f4a"
+checksum = "824a212faf96e9acacdbd09febd34438f8f711fb84e09a8916013cd7815ca28d"
 dependencies = [
  "anstyle",
  "anstyle-parse",
@@ -84,47 +78,44 @@ dependencies = [
 
 [[package]]
 name = "anstyle"
-version = "1.0.13"
+version = "1.0.14"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5192cca8006f1fd4f7237516f40fa183bb07f8fbdfedaa0036de5ea9b0b45e78"
+checksum = "940b3a0ca603d1eade50a4846a2afffd5ef57a9feac2c0e2ec2e14f9ead76000"
 
 [[package]]
 name = "anstyle-parse"
-version = "0.2.7"
+version = "1.0.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4e7644824f0aa2c7b9384579234ef10eb7efb6a0deb83f9630a49594dd9c15c2"
+checksum = "52ce7f38b242319f7cabaa6813055467063ecdc9d355bbb4ce0c68908cd8130e"
 dependencies = [
  "utf8parse",
 ]
 
 [[package]]
 name = "anstyle-query"
-version = "1.1.4"
+version = "1.1.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9e231f6134f61b71076a3eab506c379d4f36122f2af15a9ff04415ea4c3339e2"
+checksum = "40c48f72fd53cd289104fc64099abca73db4166ad86ea0b4341abe65af83dadc"
 dependencies = [
- "windows-sys 0.60.2",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
 name = "anstyle-wincon"
-version = "3.0.10"
+version = "3.0.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3e0633414522a32ffaac8ac6cc8f748e090c5717661fddeea04219e2344f5f2a"
+checksum = "291e6a250ff86cd4a820112fb8898808a366d8f9f58ce16d1f538353ad55747d"
 dependencies = [
  "anstyle",
  "once_cell_polyfill",
- "windows-sys 0.60.2",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
 name = "anyhow"
-version = "1.0.100"
+version = "1.0.102"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a23eb6b1614318a8071c9b2521f36b424b2c83db5eb3a0fead4a6c0809af6e61"
-dependencies = [
- "backtrace",
-]
+checksum = "7f202df86484c868dbad7eaa557ef785d5c66295e41b460ef922eca0723b842c"
 
 [[package]]
 name = "arbitrary"
@@ -207,7 +198,7 @@ dependencies = [
  "miniz_oxide",
  "object",
  "rustc-demangle",
- "windows-link 0.2.1",
+ "windows-link",
 ]
 
 [[package]]
@@ -259,9 +250,9 @@ checksum = "bef38d45163c2f1dde094a7dfd33ccf595c92905c8f8f4fdc18d06fb1037718a"
 
 [[package]]
 name = "bitflags"
-version = "2.10.0"
+version = "2.11.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "812e12b5285cc515a9c72a5c1d3b6d46a19dac5acfef5265968c166106e31dd3"
+checksum = "c4512299f36f043ab09a583e57bceb5a5aab7a73db1805848e8fef3c9e8c78b3"
 
 [[package]]
 name = "block"
@@ -289,15 +280,15 @@ dependencies = [
 
 [[package]]
 name = "bumpalo"
-version = "3.19.0"
+version = "3.20.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "46c5e41b57b8bba42a04676d81cb89e9ee8e859a1a66f80a5a72e1cb76b34d43"
+checksum = "5d20789868f4b01b2f2caec9f5c4e0213b41e3e5702a50157d699ae31ced2fcb"
 
 [[package]]
 name = "bytemuck"
-version = "1.24.0"
+version = "1.25.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1fbdf580320f38b612e485521afda1ee26d10cc9884efaaa750d383e13e3c5f4"
+checksum = "c8efb64bd706a16a1bdde310ae86b351e4d21550d98d056f22f8a7f7a2183fec"
 dependencies = [
  "bytemuck_derive",
 ]
@@ -327,9 +318,9 @@ checksum = "8f1fe948ff07f4bd06c30984e69f5b4899c516a3ef74f34df92a2df2ab535495"
 
 [[package]]
 name = "bytes"
-version = "1.10.1"
+version = "1.11.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d71b6127be86fdcfddb610f7182ac57211d4b18a3e9c82eb2d17662f2227ad6a"
+checksum = "1e748733b7cbc798e1434b6ac524f0c1ff2ab456fe201501e6497c8417a4fc33"
 
 [[package]]
 name = "candle-core"
@@ -342,7 +333,7 @@ dependencies = [
  "candle-kernels",
  "candle-metal-kernels",
  "candle-ug",
- "cudarc 0.19.1",
+ "cudarc 0.19.7",
  "float8 0.6.1",
  "gemm 0.19.0",
  "half",
@@ -354,27 +345,37 @@ dependencies = [
  "num_cpus",
  "objc2-foundation",
  "objc2-metal",
- "rand 0.9.2",
+ "rand 0.9.4",
  "rand_distr",
  "rayon",
  "safetensors 0.7.0",
- "thiserror 2.0.17",
- "yoke 0.8.1",
- "zip 7.4.0",
+ "thiserror 2.0.18",
+ "yoke 0.8.2",
+ "zip 7.2.0",
 ]
 
 [[package]]
 name = "candle-flash-attn"
-version = "0.9.1"
+version = "0.9.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fb38a5bfae09c4ae73fd00039e5eaf97a7d6d9400cc35ee8e603fc4a5f9cb0a3"
+checksum = "c94ddd2e7bb828777b0a8d999ed40d2d6c3c96c9ef2a3111a69e0d96efc436d2"
 dependencies = [
  "anyhow",
  "bindgen_cuda",
  "candle-core",
+ "candle-flash-attn-build",
  "half",
 ]
 
+[[package]]
+name = "candle-flash-attn-build"
+version = "0.9.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4bd79da06f2a3b831cb4f5a1ee393d6f2c5a913e28f5000c678a84108519a78c"
+dependencies = [
+ "anyhow",
+]
+
 [[package]]
 name = "candle-kernels"
 version = "0.9.2"
@@ -395,7 +396,7 @@ dependencies = [
  "objc2-foundation",
  "objc2-metal",
  "once_cell",
- "thiserror 2.0.17",
+ "thiserror 2.0.18",
  "tracing",
 ]
 
@@ -416,7 +417,7 @@ dependencies = [
  "rayon",
  "safetensors 0.7.0",
  "serde",
- "thiserror 2.0.17",
+ "thiserror 2.0.18",
 ]
 
 [[package]]
@@ -441,9 +442,9 @@ dependencies = [
 
 [[package]]
 name = "cc"
-version = "1.2.41"
+version = "1.2.62"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ac9fe6cdbb24b6ade63616c0a0688e45bb56732262c158df3c0c4bea4ca47cb7"
+checksum = "a1dce859f0832a7d088c4f1119888ab94ef4b5d6795d1ce05afb7fe159d79f98"
 dependencies = [
  "find-msvc-tools",
  "shlex",
@@ -463,23 +464,22 @@ checksum = "613afe47fcd5fac7ccf1db93babcb082c5994d996f20b8b159f2ad1658eb5724"
 
 [[package]]
 name = "chrono"
-version = "0.4.41"
+version = "0.4.44"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c469d952047f47f91b68d1cba3f10d63c11d73e4636f24f08daf0278abf01c4d"
+checksum = "c673075a2e0e5f4a1dde27ce9dee1ea4558c7ffe648f576438a20ca1d2acc4b0"
 dependencies = [
- "android-tzdata",
  "iana-time-zone",
  "js-sys",
  "num-traits",
  "wasm-bindgen",
- "windows-link 0.1.3",
+ "windows-link",
 ]
 
 [[package]]
 name = "clap"
-version = "4.5.50"
+version = "4.6.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0c2cfd7bf8a6017ddaa4e32ffe7403d547790db06bd171c1c53926faab501623"
+checksum = "1ddb117e43bbf7dacf0a4190fef4d345b9bad68dfc649cb349e7d17d28428e51"
 dependencies = [
  "clap_builder",
  "clap_derive",
@@ -487,9 +487,9 @@ dependencies = [
 
 [[package]]
 name = "clap_builder"
-version = "4.5.50"
+version = "4.6.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0a4c05b9e80c5ccd3a7ef080ad7b6ba7d6fc00a985b8b157197075677c82c7a0"
+checksum = "714a53001bf66416adb0e2ef5ac857140e7dc3a0c48fb28b2f10762fc4b5069f"
 dependencies = [
  "anstream",
  "anstyle",
@@ -499,9 +499,9 @@ dependencies = [
 
 [[package]]
 name = "clap_derive"
-version = "4.5.49"
+version = "4.6.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2a0b5487afeab2deb2ff4e03a807ad1a03ac532ff5a2cee5d86884440c7f7671"
+checksum = "f2ce8604710f6733aa641a2b3731eaa1e8b3d9973d5e3565da11800813f997a9"
 dependencies = [
  "heck",
  "proc-macro2",
@@ -511,15 +511,15 @@ dependencies = [
 
 [[package]]
 name = "clap_lex"
-version = "0.7.6"
+version = "1.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a1d728cc89cf3aee9ff92b05e62b19ee65a02b5702cff7d5a377e32c6ae29d8d"
+checksum = "c8d4a3bb8b1e0c1050499d1815f5ab16d04f0959b233085fb31653fbfc9d98f9"
 
 [[package]]
 name = "colorchoice"
-version = "1.0.4"
+version = "1.0.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b05b61dc5112cbb17e4b6cd61790d9845d13888356391624cbe7e41efeac1e75"
+checksum = "1d07550c9036bf2ae0c684c4297d503f838287c83c53686d05370d0e139ae570"
 
 [[package]]
 name = "compact_str"
@@ -549,6 +549,18 @@ dependencies = [
  "windows-sys 0.59.0",
 ]
 
+[[package]]
+name = "console"
+version = "0.16.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d64e8af5551369d19cf50138de61f1c42074ab970f74e99be916646777f8fc87"
+dependencies = [
+ "encode_unicode",
+ "libc",
+ "unicode-width",
+ "windows-sys 0.61.2",
+]
+
 [[package]]
 name = "cookie"
 version = "0.18.1"
@@ -638,9 +650,9 @@ checksum = "460fbee9c2c2f33933d720630a6a0bac33ba7053db5344fac858d4b8952d77d5"
 
 [[package]]
 name = "crypto-common"
-version = "0.1.6"
+version = "0.1.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1bfb12502f3fc46cca1bb51ac28df9d618d813cdc3d2f25b9fe775a34af26bb3"
+checksum = "78c8292055d1c1df0cce5d180393dc8cce0abec0a7102adb6c7b1eef6016d60a"
 dependencies = [
  "generic-array",
  "typenum",
@@ -658,9 +670,8 @@ dependencies = [
 
 [[package]]
 name = "cudarc"
-version = "0.19.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "60d4882b3e023670c25b8a3e1ac97349070f8e5400807c091b70e0bdad6e9b93"
+version = "0.19.7"
+source = "git+https://github.com/coreylowman/cudarc?branch=main#3e5d38b5fe5ec81c934bdc2c7207f181772e307d"
 dependencies = [
  "float8 0.7.0",
  "half",
@@ -704,9 +715,9 @@ dependencies = [
 
 [[package]]
 name = "dary_heap"
-version = "0.3.8"
+version = "0.3.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "06d2e3287df1c007e74221c49ca10a95d557349e54b3a75dc2fb14712c751f04"
+checksum = "8b1e3a325bc115f096c8b77bbf027a7c2592230e70be2d985be950d3d5e60ebe"
 dependencies = [
  "serde",
 ]
@@ -719,7 +730,7 @@ dependencies = [
  "deepseek-ocr-core",
  "deepseek-ocr-infer-deepseek",
  "hf-hub",
- "indicatif",
+ "indicatif 0.17.11",
  "once_cell",
  "reqwest 0.11.27",
  "serde",
@@ -776,7 +787,7 @@ dependencies = [
  "ndarray",
  "ndarray-npy",
  "once_cell",
- "rand 0.8.5",
+ "rand 0.8.6",
  "rayon",
  "serde",
  "serde_json",
@@ -804,7 +815,7 @@ dependencies = [
  "deepseek-ocr-dsq-models",
  "deepseek-ocr-dsq-writer",
  "half",
- "indicatif",
+ "indicatif 0.17.11",
  "memmap2",
  "rayon",
  "safetensors 0.4.5",
@@ -862,7 +873,7 @@ dependencies = [
  "ndarray",
  "ndarray-npy",
  "once_cell",
- "rand 0.8.5",
+ "rand 0.8.6",
  "rayon",
  "serde",
  "serde_json",
@@ -919,7 +930,7 @@ dependencies = [
  "memmap2",
  "ndarray",
  "ndarray-npy",
- "rand 0.8.5",
+ "rand 0.8.6",
  "safetensors 0.4.5",
  "serde",
  "serde_json",
@@ -943,7 +954,7 @@ dependencies = [
  "deepseek-ocr-infer-glm",
  "deepseek-ocr-infer-paddleocr",
  "image",
- "reqwest 0.12.24",
+ "reqwest 0.12.28",
  "rocket",
  "serde",
  "serde_json",
@@ -957,9 +968,9 @@ dependencies = [
 
 [[package]]
 name = "deranged"
-version = "0.5.4"
+version = "0.5.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a41953f86f8a05768a6cda24def994fd2f424b04ec5c719cf89989779f199071"
+checksum = "7cd812cc2bc1d69d4764bd80df88b4317eaef9e773c75226407d9bc0876b211c"
 dependencies = [
  "powerfmt",
 ]
@@ -1032,7 +1043,7 @@ version = "0.4.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b035a542cf7abf01f2e3c4d5a7acbaebfefe120ae4efc7bde3df98186e4b8af7"
 dependencies = [
- "bitflags 2.10.0",
+ "bitflags 2.11.1",
  "proc-macro2",
  "proc-macro2-diagnostics",
  "quote",
@@ -1113,16 +1124,16 @@ dependencies = [
  "libc",
  "option-ext",
  "redox_users 0.5.2",
- "windows-sys 0.60.2",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
 name = "dispatch2"
-version = "0.3.0"
+version = "0.3.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "89a09f22a6c6069a18470eb92d2298acf25463f14256d24778e1230d789a2aec"
+checksum = "1e0e367e4e7da84520dedcac1901e4da967309406d1e51017ae1abfb97adbd38"
 dependencies = [
- "bitflags 2.10.0",
+ "bitflags 2.11.1",
  "objc2",
 ]
 
@@ -1208,7 +1219,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "39cab71617ae0d63f51a36d69f866391735b51691dbda63cf6f96d042b63efeb"
 dependencies = [
  "libc",
- "windows-sys 0.60.2",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
@@ -1222,21 +1233,21 @@ dependencies = [
 
 [[package]]
 name = "fast_image_resize"
-version = "5.3.0"
+version = "5.5.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1bd1eda71e8af93f8b00e189404235d82f4de77ea4a0d182b44a7f03994d647c"
+checksum = "fbc7fe45cf92b43817ff62a3723e862b85bd1d06288f63007f7645d1d2f7a060"
 dependencies = [
  "cfg-if",
  "document-features",
  "num-traits",
- "thiserror 2.0.17",
+ "thiserror 2.0.18",
 ]
 
 [[package]]
 name = "fastrand"
-version = "2.3.0"
+version = "2.4.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "37909eebbb50d72f9059c3b6d82c0463f2ff062c9e95845c43a6c9c0355411be"
+checksum = "9f1f227452a390804cdb637b74a86990f2a7d7ba4b7d5693aac9b4dd6defd8d6"
 
 [[package]]
 name = "fdeflate"
@@ -1263,27 +1274,25 @@ dependencies = [
 
 [[package]]
 name = "filetime"
-version = "0.2.25"
+version = "0.2.29"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "35c0522e981e68cbfa8c3f978441a5f34b30b96e146b33cd3359176b50fe8586"
+checksum = "5c287a33c7f0a620c38e641e7f60827713987b3c0f26e8ddc9462cc69cf75759"
 dependencies = [
  "cfg-if",
  "libc",
- "libredox",
- "windows-sys 0.59.0",
 ]
 
 [[package]]
 name = "find-msvc-tools"
-version = "0.1.4"
+version = "0.1.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "52051878f80a721bb68ebfbc930e07b65ba72f2da88968ea5c06fd6ca3d3a127"
+checksum = "5baebc0774151f905a1a2cc41989300b1e6fbb29aff0ceffa1064fdd3088d582"
 
 [[package]]
 name = "flate2"
-version = "1.1.4"
+version = "1.1.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dc5a4e564e38c699f2880d3fda590bedc2e69f3f84cd48b457bd892ce61d0aa9"
+checksum = "843fba2746e448b37e26a819579957415c8cef339bf08564fe8b7ddbd959573c"
 dependencies = [
  "crc32fast",
  "miniz_oxide",
@@ -1295,10 +1304,10 @@ version = "0.6.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "719a903cc23e4a89e87962c2a80fdb45cdaad0983a89bd150bb57b4c8571a7d5"
 dependencies = [
- "cudarc 0.19.1",
+ "cudarc 0.19.7",
  "half",
  "num-traits",
- "rand 0.9.2",
+ "rand 0.9.4",
  "rand_distr",
 ]
 
@@ -1317,6 +1326,12 @@ version = "1.0.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "3f9eec918d3f24069decb9af1554cad7c880e2da24a9afd88aca000531ab82c1"
 
+[[package]]
+name = "foldhash"
+version = "0.1.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d9c4f5dac5e15c24eb999c26181a6ca40b39fe946cbe4c263c7209467bc83af2"
+
 [[package]]
 name = "foldhash"
 version = "0.2.0"
@@ -1361,9 +1376,9 @@ dependencies = [
 
 [[package]]
 name = "futures"
-version = "0.3.31"
+version = "0.3.32"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "65bc07b1a8bc7c85c5f2e110c476c7389b4554ba72af57d8445ea63a576b0876"
+checksum = "8b147ee9d1f6d097cef9ce628cd2ee62288d963e16fb287bd9286455b241382d"
 dependencies = [
  "futures-channel",
  "futures-core",
@@ -1375,9 +1390,9 @@ dependencies = [
 
 [[package]]
 name = "futures-channel"
-version = "0.3.31"
+version = "0.3.32"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2dff15bf788c671c1934e366d07e30c1814a8ef514e1af724a602e8a2fbe1b10"
+checksum = "07bbe89c50d7a535e539b8c17bc0b49bdb77747034daa8087407d655f3f7cc1d"
 dependencies = [
  "futures-core",
  "futures-sink",
@@ -1385,21 +1400,21 @@ dependencies = [
 
 [[package]]
 name = "futures-core"
-version = "0.3.31"
+version = "0.3.32"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "05f29059c0c2090612e8d742178b0580d2dc940c837851ad723096f87af6663e"
+checksum = "7e3450815272ef58cec6d564423f6e755e25379b217b0bc688e295ba24df6b1d"
 
 [[package]]
 name = "futures-io"
-version = "0.3.31"
+version = "0.3.32"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9e5c1b78ca4aae1ac06c48a526a655760685149f0d465d21f37abfe57ce075c6"
+checksum = "cecba35d7ad927e23624b22ad55235f2239cfa44fd10428eecbeba6d6a717718"
 
 [[package]]
 name = "futures-macro"
-version = "0.3.31"
+version = "0.3.32"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "162ee34ebcb7c64a8abebc059ce0fee27c2262618d7b60ed8faf72fef13c3650"
+checksum = "e835b70203e41293343137df5c0664546da5745f82ec9b84d40be8336958447b"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -1408,21 +1423,21 @@ dependencies = [
 
 [[package]]
 name = "futures-sink"
-version = "0.3.31"
+version = "0.3.32"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e575fab7d1e0dcb8d0c7bcf9a63ee213816ab51902e6d244a95819acacf1d4f7"
+checksum = "c39754e157331b013978ec91992bde1ac089843443c49cbc7f46150b0fad0893"
 
 [[package]]
 name = "futures-task"
-version = "0.3.31"
+version = "0.3.32"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f90f7dce0722e95104fcb095585910c0977252f286e354b5e3bd38902cd99988"
+checksum = "037711b3d59c33004d3856fbdc83b99d4ff37a24768fa1be9ce3538a1cde4393"
 
 [[package]]
 name = "futures-util"
-version = "0.3.31"
+version = "0.3.32"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9fa08315bb612088cc391249efdc3bc77536f16c91f6cf495e6fbe85b20a4a81"
+checksum = "389ca41296e6190b48053de0321d02a77f32f8a5d2461dd38762c0593805c6d6"
 dependencies = [
  "futures-channel",
  "futures-core",
@@ -1432,7 +1447,6 @@ dependencies = [
  "futures-task",
  "memchr",
  "pin-project-lite",
- "pin-utils",
  "slab",
 ]
 
@@ -1689,9 +1703,9 @@ dependencies = [
 
 [[package]]
 name = "generic-array"
-version = "0.14.9"
+version = "0.14.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4bb6743198531e02858aeaea5398fcc883e71851fcbcb5a2f773e2fb6cb1edf2"
+checksum = "85649ca51fd72272d7821adaf274ad91c288277713d9c18820d8499a7ff69e9a"
 dependencies = [
  "typenum",
  "version_check",
@@ -1699,9 +1713,9 @@ dependencies = [
 
 [[package]]
 name = "getrandom"
-version = "0.2.16"
+version = "0.2.17"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "335ff9f135e4384c8150d6f27c6daed433577f86b4750418338c01a1a2528592"
+checksum = "ff2abc00be7fca6ebc474524697ae276ad847ad0a6b3faa4bcb027e9a4614ad0"
 dependencies = [
  "cfg-if",
  "js-sys",
@@ -1719,11 +1733,24 @@ dependencies = [
  "cfg-if",
  "js-sys",
  "libc",
- "r-efi",
+ "r-efi 5.3.0",
  "wasip2",
  "wasm-bindgen",
 ]
 
+[[package]]
+name = "getrandom"
+version = "0.4.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0de51e6874e94e7bf76d726fc5d13ba782deca734ff60d5bb2fb2607c7406555"
+dependencies = [
+ "cfg-if",
+ "libc",
+ "r-efi 6.0.0",
+ "wasip2",
+ "wasip3",
+]
+
 [[package]]
 name = "getset"
 version = "0.1.6"
@@ -1777,23 +1804,39 @@ dependencies = [
  "cfg-if",
  "crunchy",
  "num-traits",
- "rand 0.9.2",
+ "rand 0.9.4",
  "rand_distr",
  "zerocopy",
 ]
 
 [[package]]
 name = "hashbrown"
-version = "0.16.0"
+version = "0.15.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5419bdc4f6a9207fbeba6d11b604d481addf78ecd10c11ad51e76c2f6482748d"
+checksum = "9229cfe53dfd69f0609a49f65461bd93001ea1ef889cd5529dd176593f5338a1"
+dependencies = [
+ "foldhash 0.1.5",
+]
+
+[[package]]
+name = "hashbrown"
+version = "0.16.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "841d1cc9bed7f9236f321df977030373f4a4163ae1a7dbfe1a51a2c1a51d9100"
 dependencies = [
  "allocator-api2",
  "equivalent",
- "foldhash",
+ "foldhash 0.2.0",
  "serde",
+ "serde_core",
 ]
 
+[[package]]
+name = "hashbrown"
+version = "0.17.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ed5909b6e89a2db4456e54cd5f673791d7eca6732202bbf2a9cc504fe2f9b84a"
+
 [[package]]
 name = "heck"
 version = "0.5.0"
@@ -1813,15 +1856,15 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "629d8f3bbeda9d148036d6b0de0a3ab947abd08ce90626327fc3547a49d59d97"
 dependencies = [
  "dirs 6.0.0",
- "http 1.3.1",
- "indicatif",
+ "http 1.4.0",
+ "indicatif 0.17.11",
  "libc",
  "log",
- "rand 0.9.2",
- "reqwest 0.12.24",
+ "rand 0.9.4",
+ "reqwest 0.12.28",
  "serde",
  "serde_json",
- "thiserror 2.0.17",
+ "thiserror 2.0.18",
  "ureq",
  "windows-sys 0.60.2",
 ]
@@ -1839,12 +1882,11 @@ dependencies = [
 
 [[package]]
 name = "http"
-version = "1.3.1"
+version = "1.4.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f4a85d31aea989eead29a3aaf9e1115a180df8282431156e533de47660892565"
+checksum = "e3ba2a386d7f85a81f119ad7498ebe444d2e22c2af0b86b069416ace48b3311a"
 dependencies = [
  "bytes",
- "fnv",
  "itoa",
 ]
 
@@ -1866,7 +1908,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1efedce1fb8e6913f23e0c92de8e62cd5b772a67e7b3946df930a62566c93184"
 dependencies = [
  "bytes",
- "http 1.3.1",
+ "http 1.4.0",
 ]
 
 [[package]]
@@ -1877,7 +1919,7 @@ checksum = "b021d93e26becf5dc7e1b75b1bed1fd93124b374ceb73f43d4d4eafec896a64a"
 dependencies = [
  "bytes",
  "futures-core",
- "http 1.3.1",
+ "http 1.4.0",
  "http-body 1.0.1",
  "pin-project-lite",
 ]
@@ -1920,20 +1962,19 @@ dependencies = [
 
 [[package]]
 name = "hyper"
-version = "1.7.0"
+version = "1.9.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "eb3aa54a13a0dfe7fbe3a59e0c76093041720fdc77b110cc0fc260fafb4dc51e"
+checksum = "6299f016b246a94207e63da54dbe807655bf9e00044f73ded42c3ac5305fbcca"
 dependencies = [
  "atomic-waker",
  "bytes",
  "futures-channel",
  "futures-core",
- "http 1.3.1",
+ "http 1.4.0",
  "http-body 1.0.1",
  "httparse",
  "itoa",
  "pin-project-lite",
- "pin-utils",
  "smallvec",
  "tokio",
  "want",
@@ -1955,40 +1996,38 @@ dependencies = [
 
 [[package]]
 name = "hyper-rustls"
-version = "0.27.7"
+version = "0.27.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e3c93eb611681b207e1fe55d5a71ecf91572ec8a6705cdb6857f7d8d5242cf58"
+checksum = "33ca68d021ef39cf6463ab54c1d0f5daf03377b70561305bb89a8f83aab66e0f"
 dependencies = [
- "http 1.3.1",
- "hyper 1.7.0",
+ "http 1.4.0",
+ "hyper 1.9.0",
  "hyper-util",
- "rustls 0.23.34",
- "rustls-pki-types",
+ "rustls 0.23.40",
  "tokio",
  "tokio-rustls 0.26.4",
  "tower-service",
- "webpki-roots 1.0.3",
+ "webpki-roots 1.0.7",
 ]
 
 [[package]]
 name = "hyper-util"
-version = "0.1.17"
+version = "0.1.20"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3c6995591a8f1380fcb4ba966a252a4b29188d51d2b89e3a252f5305be65aea8"
+checksum = "96547c2556ec9d12fb1578c4eaf448b04993e7fb79cbaad930a656880a6bdfa0"
 dependencies = [
  "base64 0.22.1",
  "bytes",
  "futures-channel",
- "futures-core",
  "futures-util",
- "http 1.3.1",
+ "http 1.4.0",
  "http-body 1.0.1",
- "hyper 1.7.0",
+ "hyper 1.9.0",
  "ipnet",
  "libc",
  "percent-encoding",
  "pin-project-lite",
- "socket2 0.6.1",
+ "socket2 0.6.3",
  "tokio",
  "tower-service",
  "tracing",
@@ -1996,9 +2035,9 @@ dependencies = [
 
 [[package]]
 name = "iana-time-zone"
-version = "0.1.63"
+version = "0.1.65"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b0c919e5debc312ad217002b8048a17b7d83f80703865bbfcfebb0458b0b27d8"
+checksum = "e31bc9ad994ba00e440a8aa5c9ef0ec67d5cb5e5cb0cc7f8b744a35b389cc470"
 dependencies = [
  "android_system_properties",
  "core-foundation-sys",
@@ -2020,22 +2059,23 @@ dependencies = [
 
 [[package]]
 name = "icu_collections"
-version = "2.0.0"
+version = "2.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "200072f5d0e3614556f94a9930d5dc3e0662a652823904c3a75dc3b0af7fee47"
+checksum = "2984d1cd16c883d7935b9e07e44071dca8d917fd52ecc02c04d5fa0b5a3f191c"
 dependencies = [
  "displaydoc",
  "potential_utf",
- "yoke 0.8.1",
+ "utf8_iter",
+ "yoke 0.8.2",
  "zerofrom",
  "zerovec",
 ]
 
 [[package]]
 name = "icu_locale_core"
-version = "2.0.0"
+version = "2.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0cde2700ccaed3872079a65fb1a78f6c0a36c91570f28755dda67bc8f7d9f00a"
+checksum = "92219b62b3e2b4d88ac5119f8904c10f8f61bf7e95b640d25ba3075e6cac2c29"
 dependencies = [
  "displaydoc",
  "litemap",
@@ -2046,11 +2086,10 @@ dependencies = [
 
 [[package]]
 name = "icu_normalizer"
-version = "2.0.0"
+version = "2.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "436880e8e18df4d7bbc06d58432329d6458cc84531f7ac5f024e93deadb37979"
+checksum = "c56e5ee99d6e3d33bd91c5d85458b6005a22140021cc324cea84dd0e72cff3b4"
 dependencies = [
- "displaydoc",
  "icu_collections",
  "icu_normalizer_data",
  "icu_properties",
@@ -2061,49 +2100,51 @@ dependencies = [
 
 [[package]]
 name = "icu_normalizer_data"
-version = "2.0.0"
+version = "2.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "00210d6893afc98edb752b664b8890f0ef174c8adbb8d0be9710fa66fbbf72d3"
+checksum = "da3be0ae77ea334f4da67c12f149704f19f81d1adf7c51cf482943e84a2bad38"
 
 [[package]]
 name = "icu_properties"
-version = "2.0.1"
+version = "2.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "016c619c1eeb94efb86809b015c58f479963de65bdb6253345c1a1276f22e32b"
+checksum = "bee3b67d0ea5c2cca5003417989af8996f8604e34fb9ddf96208a033901e70de"
 dependencies = [
- "displaydoc",
  "icu_collections",
  "icu_locale_core",
  "icu_properties_data",
  "icu_provider",
- "potential_utf",
  "zerotrie",
  "zerovec",
 ]
 
 [[package]]
 name = "icu_properties_data"
-version = "2.0.1"
+version = "2.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "298459143998310acd25ffe6810ed544932242d3f07083eee1084d83a71bd632"
+checksum = "8e2bbb201e0c04f7b4b3e14382af113e17ba4f63e2c9d2ee626b720cbce54a14"
 
 [[package]]
 name = "icu_provider"
-version = "2.0.0"
+version = "2.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "03c80da27b5f4187909049ee2d72f276f0d9f99a42c306bd0131ecfe04d8e5af"
+checksum = "139c4cf31c8b5f33d7e199446eff9c1e02decfc2f0eec2c8d71f65befa45b421"
 dependencies = [
  "displaydoc",
  "icu_locale_core",
- "stable_deref_trait",
- "tinystr",
  "writeable",
- "yoke 0.8.1",
+ "yoke 0.8.2",
  "zerofrom",
  "zerotrie",
  "zerovec",
 ]
 
+[[package]]
+name = "id-arena"
+version = "2.3.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3d3067d79b975e8844ca9eb072e16b31c3c1c36928edf9c6789548c524d0d954"
+
 [[package]]
 name = "ident_case"
 version = "1.0.1"
@@ -2123,9 +2164,9 @@ dependencies = [
 
 [[package]]
 name = "idna_adapter"
-version = "1.2.1"
+version = "1.2.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3acae9609540aa318d1bc588455225fb2085b9ed0c4f6bd0d9d5bcd86f1a0344"
+checksum = "cb68373c0d6620ef8105e855e7745e18b0d00d3bdb07fb532e434244cdb9a714"
 dependencies = [
  "icu_normalizer",
  "icu_properties",
@@ -2133,9 +2174,9 @@ dependencies = [
 
 [[package]]
 name = "image"
-version = "0.25.8"
+version = "0.25.10"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "529feb3e6769d234375c4cf1ee2ce713682b8e76538cb13f9fc23e1400a591e7"
+checksum = "85ab80394333c02fe689eaf900ab500fbd0c2213da414687ebf995a65d5a6104"
 dependencies = [
  "bytemuck",
  "byteorder-lite",
@@ -2148,12 +2189,12 @@ dependencies = [
 
 [[package]]
 name = "indexmap"
-version = "2.12.0"
+version = "2.14.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6717a8d2a5a929a1a2eb43a12812498ed141a0bcfb7e8f7844fbdbe4303bba9f"
+checksum = "d466e9454f08e4a911e14806c24e16fba1b4c121d1ea474396f396069cf949d9"
 dependencies = [
  "equivalent",
- "hashbrown",
+ "hashbrown 0.17.1",
  "serde",
  "serde_core",
 ]
@@ -2164,13 +2205,26 @@ version = "0.17.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "183b3088984b400f4cfac3620d5e076c84da5364016b4f49473de574b2586235"
 dependencies = [
- "console",
+ "console 0.15.11",
  "number_prefix",
  "portable-atomic",
  "unicode-width",
  "web-time",
 ]
 
+[[package]]
+name = "indicatif"
+version = "0.18.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "25470f23803092da7d239834776d653104d551bc4d7eacaf31e6837854b8e9eb"
+dependencies = [
+ "console 0.16.3",
+ "portable-atomic",
+ "unicode-width",
+ "unit-prefix",
+ "web-time",
+]
+
 [[package]]
 name = "inlinable_string"
 version = "0.1.15"
@@ -2201,19 +2255,9 @@ dependencies = [
 
 [[package]]
 name = "ipnet"
-version = "2.11.0"
+version = "2.12.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "469fb0b9cefa57e3ef31275ee7cacb78f2fdca44e4765491884a2b119d4eb130"
-
-[[package]]
-name = "iri-string"
-version = "0.7.8"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dbc5ebe9c3a1a7a5127f920a418f7585e9e758e911d0466ed004f393b0e380b2"
-dependencies = [
- "memchr",
- "serde",
-]
+checksum = "d98f6fed1fde3f8c21bc40a1abb88dd75e67924f9cffc3ef95607bad8017f8e2"
 
 [[package]]
 name = "is-terminal"
@@ -2223,7 +2267,7 @@ checksum = "3640c1c38b8e4e43584d8df18be5fc6b0aa314ce6ebf51b53313d4306cca8e46"
 dependencies = [
  "hermit-abi",
  "libc",
- "windows-sys 0.60.2",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
@@ -2243,16 +2287,18 @@ dependencies = [
 
 [[package]]
 name = "itoa"
-version = "1.0.15"
+version = "1.0.18"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4a5f13b858c8d314ee3e8f639011f7ccefe71f97f96e50151fb991f267928e2c"
+checksum = "8f42a60cbdf9a97f5d2305f08a87dc4e09308d1276d28c869c684d7777685682"
 
 [[package]]
 name = "js-sys"
-version = "0.3.81"
+version = "0.3.98"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ec48937a97411dcb524a265206ccd4c90bb711fca92b2792c407f268825b9305"
+checksum = "67df7112613f8bfd9150013a0314e196f4800d3201ae742489d999db2f979f08"
 dependencies = [
+ "cfg-if",
+ "futures-util",
  "once_cell",
  "wasm-bindgen",
 ]
@@ -2263,11 +2309,17 @@ version = "1.5.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "bbd2bcb4c963f2ddae06a2efc7e9f3591312473c50c6685e1f298068316e66fe"
 
+[[package]]
+name = "leb128fmt"
+version = "0.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "09edd9e8b54e49e587e4f6295a7d29c3ea94d469cb40ab8ca70b288248a81db2"
+
 [[package]]
 name = "libc"
-version = "0.2.177"
+version = "0.2.186"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2874a2af47a2325c2001a6e6fad9b16a53b802102b528163885171cf92b15976"
+checksum = "68ab91017fe16c622486840e4c83c9a37afeff978bd239b5293d61ece587de66"
 
 [[package]]
 name = "libloading"
@@ -2276,7 +2328,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "d7c4b02199fee7c5d21a5ae7d8cfa79a6ef5bb2fc834d6e9058e89c825efdc55"
 dependencies = [
  "cfg-if",
- "windows-link 0.2.1",
+ "windows-link",
 ]
 
 [[package]]
@@ -2286,37 +2338,35 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "754ca22de805bb5744484a5b151a9e1a8e837d5dc232c2d7d8c2e3492edc8b60"
 dependencies = [
  "cfg-if",
- "windows-link 0.2.1",
+ "windows-link",
 ]
 
 [[package]]
 name = "libm"
-version = "0.2.15"
+version = "0.2.16"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f9fbbcab51052fe104eb5e5d351cf728d30a5be1fe14d9be8a3b097481fb97de"
+checksum = "b6d2cec3eae94f9f509c767b45932f1ada8350c4bdb85af2fcab4a3c14807981"
 
 [[package]]
 name = "libredox"
-version = "0.1.10"
+version = "0.1.16"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "416f7e718bdb06000964960ffa43b4335ad4012ae8b99060261aa4a8088d5ccb"
+checksum = "e02f3bb43d335493c96bf3fd3a321600bf6bd07ed34bc64118e9293bdffea46c"
 dependencies = [
- "bitflags 2.10.0",
  "libc",
- "redox_syscall",
 ]
 
 [[package]]
 name = "linux-raw-sys"
-version = "0.11.0"
+version = "0.12.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "df1d3c3b53da64cf5760482273a98e575c651a67eec7f77df96b5b642de8f039"
+checksum = "32a66949e030da00e8c7d4434b251670a91556f4144941d37452769c25d58a53"
 
 [[package]]
 name = "litemap"
-version = "0.8.0"
+version = "0.8.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "241eaef5fd12c88705a01fc1066c48c4b36e0dd4377dcdc7ec3942cea7a69956"
+checksum = "92daf443525c4cce67b150400bc2316076100ce0b3686209eb8cf3c31612e6f0"
 
 [[package]]
 name = "litrs"
@@ -2335,9 +2385,9 @@ dependencies = [
 
 [[package]]
 name = "log"
-version = "0.4.28"
+version = "0.4.29"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "34080505efa8e45a4b816c349525ebe327ceaa8559756f0356cba97ef3bf7432"
+checksum = "5e5032e24019045c762d3c0f28f5b6b8bbf38563a65908389bf7978758920897"
 
 [[package]]
 name = "loom"
@@ -2406,15 +2456,15 @@ dependencies = [
 
 [[package]]
 name = "memchr"
-version = "2.7.6"
+version = "2.8.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f52b00d39961fc5b2736ea853c9cc86238e165017a493d1d5c8eac6bdc4cc273"
+checksum = "f8ca58f447f06ed17d5fc4043ce1b10dd205e060fb3ce5b979b8ed8e59ff3f79"
 
 [[package]]
 name = "memmap2"
-version = "0.9.8"
+version = "0.9.10"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "843a98750cd611cc2965a8213b53b43e715f13c37a9e096c6408e69990961db7"
+checksum = "714098028fe011992e1c3962653c96b2d578c4b4bce9036e15ff220319b1e0e3"
 dependencies = [
  "libc",
  "stable_deref_trait",
@@ -2426,7 +2476,7 @@ version = "0.29.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7ecfd3296f8c56b7c1f6fbac3c71cefa9d78ce009850c45000015f206dc7fa21"
 dependencies = [
- "bitflags 2.10.0",
+ "bitflags 2.11.1",
  "block",
  "core-graphics-types",
  "foreign-types",
@@ -2465,9 +2515,9 @@ checksum = "c505b3e17ed6b70a7ed2e67fbb2c560ee327353556120d6e72f5232b6880d536"
 
 [[package]]
 name = "mio"
-version = "1.1.0"
+version = "1.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "69d83b0086dc8ecf3ce9ae2874b2d1290252e2a30720bea58a5c6639b0092873"
+checksum = "50b7e5b27aa02a74bac8c3f23f448f8d87ff11f92d3aac1a6ed369ee08cc56c1"
 dependencies = [
  "libc",
  "wasi",
@@ -2498,9 +2548,9 @@ dependencies = [
 
 [[package]]
 name = "moxcms"
-version = "0.7.7"
+version = "0.8.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c588e11a3082784af229e23e8e4ecf5bcc6fbe4f69101e0421ce8d79da7f0b40"
+checksum = "bb85c154ba489f01b25c0d36ae69a87e4a1c73a72631fc6c0eb6dde34a73e44b"
 dependencies = [
  "num-traits",
  "pxfm",
@@ -2515,7 +2565,7 @@ dependencies = [
  "bytes",
  "encoding_rs",
  "futures-util",
- "http 1.3.1",
+ "http 1.4.0",
  "httparse",
  "memchr",
  "mime",
@@ -2610,9 +2660,9 @@ dependencies = [
 
 [[package]]
 name = "num-conv"
-version = "0.1.0"
+version = "0.2.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "51d515d32fb182ee37cda2ccdcb92950d6a3c2893aa280e540671c2cd0f3b1d9"
+checksum = "c6673768db2d862beb9b39a78fdcb1a69439615d5794a1be50caa9bc92c81967"
 
 [[package]]
 name = "num-integer"
@@ -2682,9 +2732,9 @@ dependencies = [
 
 [[package]]
 name = "objc2"
-version = "0.6.3"
+version = "0.6.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b7c2599ce0ec54857b29ce62166b0ed9b4f6f1a70ccc9a71165b6154caca8c05"
+checksum = "3a12a8ed07aefc768292f076dc3ac8c48f3781c8f2d5851dd3d98950e8c5a89f"
 dependencies = [
  "objc2-encode",
 ]
@@ -2695,7 +2745,7 @@ version = "0.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "2a180dd8642fa45cdb7dd721cd4c11b1cadd4929ce112ebd8b9f5803cc79d536"
 dependencies = [
- "bitflags 2.10.0",
+ "bitflags 2.11.1",
  "dispatch2",
  "objc2",
 ]
@@ -2712,7 +2762,7 @@ version = "0.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "e3e0adef53c21f888deb4fa59fc59f7eb17404926ee8a6f59f5df0fd7f9f3272"
 dependencies = [
- "bitflags 2.10.0",
+ "bitflags 2.11.1",
  "block2",
  "libc",
  "objc2",
@@ -2725,7 +2775,7 @@ version = "0.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "a0125f776a10d00af4152d74616409f0d4a2053a6f57fa5b7d6aa2854ac04794"
 dependencies = [
- "bitflags 2.10.0",
+ "bitflags 2.11.1",
  "block2",
  "dispatch2",
  "objc2",
@@ -2786,9 +2836,9 @@ dependencies = [
 
 [[package]]
 name = "once_cell"
-version = "1.21.3"
+version = "1.21.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "42f5e15c9953c5e4ccceeb2e7382a716482c34515315f7b03532b8b4e8393d2d"
+checksum = "9f7c3e4beb33f85d45ae3e3a1792185706c8e16d043238c593331cc7cd313b50"
 
 [[package]]
 name = "once_cell_polyfill"
@@ -2798,11 +2848,11 @@ checksum = "384b8ab6d37215f3c5301a95a4accb5d64aa607f1fcb26a11b5303878451b4fe"
 
 [[package]]
 name = "onig"
-version = "6.5.1"
+version = "6.5.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "336b9c63443aceef14bea841b899035ae3abe89b7c486aaf4c5bd8aafedac3f0"
+checksum = "0cc3cbf698f9438986c11a880c90a6d04b9de27575afd28bbf45b154b6c709e2"
 dependencies = [
- "bitflags 2.10.0",
+ "bitflags 2.11.1",
  "libc",
  "once_cell",
  "onig_sys",
@@ -2810,9 +2860,9 @@ dependencies = [
 
 [[package]]
 name = "onig_sys"
-version = "69.9.1"
+version = "69.9.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c7f86c6eef3d6df15f23bcfb6af487cbd2fed4e5581d58d5bf1f5f8b7f6727dc"
+checksum = "1e68317604e77e53b85896388e1a803c1d21b74c899ec9e5e1112db90735edd7"
 dependencies = [
  "cc",
  "pkg-config",
@@ -2850,7 +2900,7 @@ dependencies = [
  "libc",
  "redox_syscall",
  "smallvec",
- "windows-link 0.2.1",
+ "windows-link",
 ]
 
 [[package]]
@@ -2890,9 +2940,9 @@ checksum = "9b4f627cb1b25917193a259e49bdad08f671f8d9708acfd5fe0a8c1455d87220"
 
 [[package]]
 name = "pest"
-version = "2.8.3"
+version = "2.8.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "989e7521a040efde50c3ab6bbadafbe15ab6dc042686926be59ac35d74607df4"
+checksum = "e0848c601009d37dfa3430c4666e147e49cdcf1b92ecd3e63657d8a5f19da662"
 dependencies = [
  "memchr",
  "ucd-trie",
@@ -2900,9 +2950,9 @@ dependencies = [
 
 [[package]]
 name = "pest_derive"
-version = "2.8.3"
+version = "2.8.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "187da9a3030dbafabbbfb20cb323b976dc7b7ce91fcd84f2f74d6e31d378e2de"
+checksum = "11f486f1ea21e6c10ed15d5a7c77165d0ee443402f0780849d1768e7d9d6fe77"
 dependencies = [
  "pest",
  "pest_generator",
@@ -2910,9 +2960,9 @@ dependencies = [
 
 [[package]]
 name = "pest_generator"
-version = "2.8.3"
+version = "2.8.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "49b401d98f5757ebe97a26085998d6c0eecec4995cad6ab7fc30ffdf4b052843"
+checksum = "8040c4647b13b210a963c1ed407c1ff4fdfa01c31d6d2a098218702e6664f94f"
 dependencies = [
  "pest",
  "pest_meta",
@@ -2923,9 +2973,9 @@ dependencies = [
 
 [[package]]
 name = "pest_meta"
-version = "2.8.3"
+version = "2.8.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "72f27a2cfee9f9039c4d86faa5af122a0ac3851441a34865b8a043b46be0065a"
+checksum = "89815c69d36021a140146f26659a81d6c2afa33d216d736dd4be5381a7362220"
 dependencies = [
  "pest",
  "sha2",
@@ -2933,29 +2983,23 @@ dependencies = [
 
 [[package]]
 name = "pin-project-lite"
-version = "0.2.16"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3b3cff922bd51709b605d9ead9aa71031d81447142d828eb4a6eba76fe619f9b"
-
-[[package]]
-name = "pin-utils"
-version = "0.1.0"
+version = "0.2.17"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8b870d8c151b6f2fb93e84a13146138f05d02ed11c7e7c54f8826aaaf7c9f184"
+checksum = "a89322df9ebe1c1578d689c92318e070967d1042b512afbe49518723f4e6d5cd"
 
 [[package]]
 name = "pkg-config"
-version = "0.3.32"
+version = "0.3.33"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7edddbd0b52d732b21ad9a5fab5c704c14cd949e5e9a1ec5929a24fded1b904c"
+checksum = "19f132c84eca552bf34cab8ec81f1c1dcc229b811638f9d283dceabe58c5569e"
 
 [[package]]
 name = "png"
-version = "0.18.0"
+version = "0.18.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "97baced388464909d42d89643fe4361939af9b7ce7a31ee32a168f832a70f2a0"
+checksum = "60769b8b31b2a9f263dae2776c37b1b28ae246943cf719eb6946a1db05128a61"
 dependencies = [
- "bitflags 2.10.0",
+ "bitflags 2.11.1",
  "crc32fast",
  "fdeflate",
  "flate2",
@@ -2964,24 +3008,24 @@ dependencies = [
 
 [[package]]
 name = "portable-atomic"
-version = "1.11.1"
+version = "1.13.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f84267b20a16ea918e43c6a88433c2d54fa145c92a811b5b047ccbe153674483"
+checksum = "c33a9471896f1c69cecef8d20cbe2f7accd12527ce60845ff44c153bb2a21b49"
 
 [[package]]
 name = "portable-atomic-util"
-version = "0.2.4"
+version = "0.2.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d8a2f0d8d040d7848a709caf78912debcc3f33ee4b3cac47d73d1e1069e83507"
+checksum = "c2a106d1259c23fac8e543272398ae0e3c0b8d33c88ed73d0cc71b0f1d902618"
 dependencies = [
  "portable-atomic",
 ]
 
 [[package]]
 name = "potential_utf"
-version = "0.1.3"
+version = "0.1.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "84df19adbe5b5a0782edcab45899906947ab039ccf4573713735ee7de1e6b08a"
+checksum = "0103b1cef7ec0cf76490e969665504990193874ea05c85ff9bab8b911d0a0564"
 dependencies = [
  "zerovec",
 ]
@@ -3001,6 +3045,16 @@ dependencies = [
  "zerocopy",
 ]
 
+[[package]]
+name = "prettyplease"
+version = "0.2.37"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "479ca8adacdd7ce8f1fb39ce9ecccbfe93a3f1344b3d0d97f20bc0196208f62b"
+dependencies = [
+ "proc-macro2",
+ "syn",
+]
+
 [[package]]
 name = "proc-macro-error-attr2"
 version = "2.0.0"
@@ -3025,9 +3079,9 @@ dependencies = [
 
 [[package]]
 name = "proc-macro2"
-version = "1.0.101"
+version = "1.0.106"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "89ae43fd86e4158d6db51ad8e2b80f313af9cc74f5c0e03ccb87de09998732de"
+checksum = "8fd00f0bb2e90d81d1044c2b32617f68fcb9fa3bb7640c23e9c748e53fb30934"
 dependencies = [
  "unicode-ident",
 ]
@@ -3084,12 +3138,9 @@ checksum = "40e24eee682d89fb193496edf918a7f407d30175b2e785fe057e4392dfd182e0"
 
 [[package]]
 name = "pxfm"
-version = "0.1.25"
+version = "0.1.29"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a3cbdf373972bf78df4d3b518d07003938e2c7d1fb5891e55f9cb6df57009d84"
-dependencies = [
- "num-traits",
-]
+checksum = "e0c5ccf5294c6ccd63a74f1565028353830a9c2f5eb0c682c355c471726a6e3f"
 
 [[package]]
 name = "py_literal"
@@ -3115,10 +3166,10 @@ dependencies = [
  "pin-project-lite",
  "quinn-proto",
  "quinn-udp",
- "rustc-hash 2.1.1",
- "rustls 0.23.34",
- "socket2 0.6.1",
- "thiserror 2.0.17",
+ "rustc-hash 2.1.2",
+ "rustls 0.23.40",
+ "socket2 0.6.3",
+ "thiserror 2.0.18",
  "tokio",
  "tracing",
  "web-time",
@@ -3126,20 +3177,20 @@ dependencies = [
 
 [[package]]
 name = "quinn-proto"
-version = "0.11.13"
+version = "0.11.14"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f1906b49b0c3bc04b5fe5d86a77925ae6524a19b816ae38ce1e426255f1d8a31"
+checksum = "434b42fec591c96ef50e21e886936e66d3cc3f737104fdb9b737c40ffb94c098"
 dependencies = [
  "bytes",
  "getrandom 0.3.4",
  "lru-slab",
- "rand 0.9.2",
+ "rand 0.9.4",
  "ring",
- "rustc-hash 2.1.1",
- "rustls 0.23.34",
+ "rustc-hash 2.1.2",
+ "rustls 0.23.40",
  "rustls-pki-types",
  "slab",
- "thiserror 2.0.17",
+ "thiserror 2.0.18",
  "tinyvec",
  "tracing",
  "web-time",
@@ -3154,16 +3205,16 @@ dependencies = [
  "cfg_aliases",
  "libc",
  "once_cell",
- "socket2 0.6.1",
+ "socket2 0.6.3",
  "tracing",
  "windows-sys 0.60.2",
 ]
 
 [[package]]
 name = "quote"
-version = "1.0.41"
+version = "1.0.45"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ce25767e7b499d1b604768e7cde645d14cc8584231ea6b295e9c9eb22c02e1d1"
+checksum = "41f2619966050689382d2b44f664f4bc593e129785a36d6ee376ddf37259b924"
 dependencies = [
  "proc-macro2",
 ]
@@ -3174,11 +3225,17 @@ version = "5.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "69cdb34c158ceb288df11e18b4bd39de994f6657d83847bdffdbd7f346754b0f"
 
+[[package]]
+name = "r-efi"
+version = "6.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f8dcc9c7d52a811697d2151c701e0d08956f92b0e24136cf4cf27b57a6a0d9bf"
+
 [[package]]
 name = "rand"
-version = "0.8.5"
+version = "0.8.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "34af8d1a0e25924bc5b7c43c079c942339d8f0a8b57c39049bef581b46327404"
+checksum = "5ca0ecfa931c29007047d1bc58e623ab12e5590e8c7cc53200d5202b69266d8a"
 dependencies = [
  "libc",
  "rand_chacha 0.3.1",
@@ -3187,12 +3244,12 @@ dependencies = [
 
 [[package]]
 name = "rand"
-version = "0.9.2"
+version = "0.9.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6db2770f06117d490610c7488547d543617b21bfa07796d7a12f6f1bd53850d1"
+checksum = "44c5af06bb1b7d3216d91932aed5265164bf384dc89cd6ba05cf59a35f5f76ea"
 dependencies = [
  "rand_chacha 0.9.0",
- "rand_core 0.9.3",
+ "rand_core 0.9.5",
 ]
 
 [[package]]
@@ -3212,7 +3269,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "d3022b5f1df60f26e1ffddd6c66e8aa15de382ae63b3a0c1bfc0e4d3e3f325cb"
 dependencies = [
  "ppv-lite86",
- "rand_core 0.9.3",
+ "rand_core 0.9.5",
 ]
 
 [[package]]
@@ -3221,14 +3278,14 @@ version = "0.6.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ec0be4795e2f6a28069bec0b5ff3e2ac9bafc99e6a9a7dc3547996c5c816922c"
 dependencies = [
- "getrandom 0.2.16",
+ "getrandom 0.2.17",
 ]
 
 [[package]]
 name = "rand_core"
-version = "0.9.3"
+version = "0.9.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "99d9a13982dcf210057a8a78572b2217b667c3beacbf3a0d8b454f6f82837d38"
+checksum = "76afc826de14238e6e8c374ddcc1fa19e374fd8dd986b0d2af0d02377261d83c"
 dependencies = [
  "getrandom 0.3.4",
 ]
@@ -3240,7 +3297,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6a8615d50dcf34fa31f7ab52692afec947c4dd0ab803cc87cb3b0b4570ff7463"
 dependencies = [
  "num-traits",
- "rand 0.9.2",
+ "rand 0.9.4",
 ]
 
 [[package]]
@@ -3249,7 +3306,7 @@ version = "11.6.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "498cd0dc59d73224351ee52a95fee0f1a617a2eae0e7d9d720cc622c73a54186"
 dependencies = [
- "bitflags 2.10.0",
+ "bitflags 2.11.1",
 ]
 
 [[package]]
@@ -3260,9 +3317,9 @@ checksum = "60a357793950651c4ed0f3f52338f53b2f809f32d83a07f72909fa13e4c6c1e3"
 
 [[package]]
 name = "rayon"
-version = "1.11.0"
+version = "1.12.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "368f01d005bf8fd9b1206fb6fa653e6c4a81ceb1466406b81792d87c5677a58f"
+checksum = "fb39b166781f92d482534ef4b4b1b2568f42613b53e5b6c160e24cfbfa30926d"
 dependencies = [
  "either",
  "rayon-core",
@@ -3301,7 +3358,7 @@ version = "0.5.18"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ed2bf2547551a7053d6fdfafda3f938979645c44812fbfcda098faae3f1a362d"
 dependencies = [
- "bitflags 2.10.0",
+ "bitflags 2.11.1",
 ]
 
 [[package]]
@@ -3310,7 +3367,7 @@ version = "0.4.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ba009ff324d1fc1b900bd1fdb31564febe58a8ccc8a6fdbb93b543d33b13ca43"
 dependencies = [
- "getrandom 0.2.16",
+ "getrandom 0.2.17",
  "libredox",
  "thiserror 1.0.69",
 ]
@@ -3321,9 +3378,9 @@ version = "0.5.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "a4e608c6638b9c18977b00b475ac1f28d14e84b27d8d42f70e0bf1e3dec127ac"
 dependencies = [
- "getrandom 0.2.16",
+ "getrandom 0.2.17",
  "libredox",
- "thiserror 2.0.17",
+ "thiserror 2.0.18",
 ]
 
 [[package]]
@@ -3348,14 +3405,14 @@ dependencies = [
 
 [[package]]
 name = "regex"
-version = "1.12.2"
+version = "1.12.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "843bc0191f75f3e22651ae5f1e72939ab2f72a4bc30fa80a066bd66edefc24d4"
+checksum = "e10754a14b9137dd7b1e3e5b0493cc9171fdd105e0ab477f51b72e7f3ac0e276"
 dependencies = [
  "aho-corasick",
  "memchr",
- "regex-automata 0.4.13",
- "regex-syntax 0.8.8",
+ "regex-automata 0.4.14",
+ "regex-syntax 0.8.10",
 ]
 
 [[package]]
@@ -3369,13 +3426,13 @@ dependencies = [
 
 [[package]]
 name = "regex-automata"
-version = "0.4.13"
+version = "0.4.14"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5276caf25ac86c8d810222b3dbb938e512c55c6831a10f3e6ed1c93b84041f1c"
+checksum = "6e1dd4122fc1595e8162618945476892eefca7b88c52820e74af6262213cae8f"
 dependencies = [
  "aho-corasick",
  "memchr",
- "regex-syntax 0.8.8",
+ "regex-syntax 0.8.10",
 ]
 
 [[package]]
@@ -3386,9 +3443,9 @@ checksum = "f162c6dd7b008981e4d40210aca20b4bd0f9b60ca9271061b07f78537722f2e1"
 
 [[package]]
 name = "regex-syntax"
-version = "0.8.8"
+version = "0.8.10"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7a2d987857b319362043e95f5353c0535c1f58eec5336fdfcf626430af7def58"
+checksum = "dc897dd8d9e8bd1ed8cdad82b5966c3e0ecae09fb1907d58efaa013543185d0a"
 
 [[package]]
 name = "reqwest"
@@ -3433,27 +3490,27 @@ dependencies = [
 
 [[package]]
 name = "reqwest"
-version = "0.12.24"
+version = "0.12.28"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9d0946410b9f7b082a427e4ef5c8ff541a88b357bc6c637c40db3a68ac70a36f"
+checksum = "eddd3ca559203180a307f12d114c268abf583f59b03cb906fd0b3ff8646c1147"
 dependencies = [
  "base64 0.22.1",
  "bytes",
  "futures-channel",
  "futures-core",
  "futures-util",
- "http 1.3.1",
+ "http 1.4.0",
  "http-body 1.0.1",
  "http-body-util",
- "hyper 1.7.0",
- "hyper-rustls 0.27.7",
+ "hyper 1.9.0",
+ "hyper-rustls 0.27.9",
  "hyper-util",
  "js-sys",
  "log",
  "percent-encoding",
  "pin-project-lite",
  "quinn",
- "rustls 0.23.34",
+ "rustls 0.23.40",
  "rustls-pki-types",
  "serde",
  "serde_json",
@@ -3470,7 +3527,7 @@ dependencies = [
  "wasm-bindgen-futures",
  "wasm-streams",
  "web-sys",
- "webpki-roots 1.0.3",
+ "webpki-roots 1.0.7",
 ]
 
 [[package]]
@@ -3481,7 +3538,7 @@ checksum = "a4689e6c2294d81e88dc6261c768b63bc4fcdb852be6d1352498b114f61383b7"
 dependencies = [
  "cc",
  "cfg-if",
- "getrandom 0.2.16",
+ "getrandom 0.2.17",
  "libc",
  "untrusted",
  "windows-sys 0.52.0",
@@ -3508,7 +3565,7 @@ dependencies = [
  "num_cpus",
  "parking_lot",
  "pin-project-lite",
- "rand 0.8.5",
+ "rand 0.8.6",
  "ref-cast",
  "rocket_codegen",
  "rocket_http",
@@ -3571,9 +3628,9 @@ dependencies = [
 
 [[package]]
 name = "rustc-demangle"
-version = "0.1.26"
+version = "0.1.27"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "56f7d92ca342cea22a06f2121d944b4fd82af56988c270852495420f961d4ace"
+checksum = "b50b8869d9fc858ce7266cce0194bd74df58b9d0e3f6df3a9fc8eb470d95c09d"
 
 [[package]]
 name = "rustc-hash"
@@ -3583,21 +3640,21 @@ checksum = "08d43f7aa6b08d49f382cde6a7982047c3426db949b1424bc4b7ec9ae12c6ce2"
 
 [[package]]
 name = "rustc-hash"
-version = "2.1.1"
+version = "2.1.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "357703d41365b4b27c590e3ed91eabb1b663f07c4c084095e60cbed4362dff0d"
+checksum = "94300abf3f1ae2e2b8ffb7b58043de3d399c73fa6f4b73826402a5c457614dbe"
 
 [[package]]
 name = "rustix"
-version = "1.1.2"
+version = "1.1.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cd15f8a2c5551a84d56efdc1cd049089e409ac19a3072d5037a17fd70719ff3e"
+checksum = "b6fe4565b9518b83ef4f91bb47ce29620ca828bd32cb7e408f0062e9930ba190"
 dependencies = [
- "bitflags 2.10.0",
+ "bitflags 2.11.1",
  "errno",
  "libc",
  "linux-raw-sys",
- "windows-sys 0.60.2",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
@@ -3614,15 +3671,15 @@ dependencies = [
 
 [[package]]
 name = "rustls"
-version = "0.23.34"
+version = "0.23.40"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6a9586e9ee2b4f8fab52a0048ca7334d7024eef48e2cb9407e3497bb7cab7fa7"
+checksum = "ef86cd5876211988985292b91c96a8f2d298df24e75989a43a3c73f2d4d8168b"
 dependencies = [
  "log",
  "once_cell",
  "ring",
  "rustls-pki-types",
- "rustls-webpki 0.103.7",
+ "rustls-webpki 0.103.13",
  "subtle",
  "zeroize",
 ]
@@ -3638,9 +3695,9 @@ dependencies = [
 
 [[package]]
 name = "rustls-pki-types"
-version = "1.12.0"
+version = "1.14.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "229a4a4c221013e7e1f1a043678c5cc39fe5171437c88fb47151a21e6f5b5c79"
+checksum = "30a7197ae7eb376e574fe940d068c30fe0462554a3ddbe4eca7838e049c937a9"
 dependencies = [
  "web-time",
  "zeroize",
@@ -3658,9 +3715,9 @@ dependencies = [
 
 [[package]]
 name = "rustls-webpki"
-version = "0.103.7"
+version = "0.103.13"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e10b3f4191e8a80e6b43eebabfac91e5dcecebb27a71f04e820c47ec41d314bf"
+checksum = "61c429a8649f110dddef65e2a5ad240f747e85f7758a6bccc7e5777bd33f756e"
 dependencies = [
  "ring",
  "rustls-pki-types",
@@ -3675,9 +3732,9 @@ checksum = "b39cdef0fa800fc44525c84ccb54a029961a8215f9619753635a9c0d2538d46d"
 
 [[package]]
 name = "ryu"
-version = "1.0.20"
+version = "1.0.23"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "28d3b2b1366ec20994f1fd18c3c594f05c5dd4bc44d8bb0c1c632c8d6829481f"
+checksum = "9774ba4a74de5f7b1c1451ed6cd5285a32eddb5cccb8cc655a4e50009e06477f"
 
 [[package]]
 name = "safetensors"
@@ -3695,7 +3752,7 @@ version = "0.7.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "675656c1eabb620b921efea4f9199f97fc86e36dd6ffd1fbbe48d0f59a4987f5"
 dependencies = [
- "hashbrown",
+ "hashbrown 0.16.1",
  "serde",
  "serde_json",
 ]
@@ -3731,6 +3788,12 @@ dependencies = [
  "untrusted",
 ]
 
+[[package]]
+name = "semver"
+version = "1.0.28"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8a7852d02fc848982e0c167ef163aaff9cd91dc640ba85e263cb1ce46fae51cd"
+
 [[package]]
 name = "seq-macro"
 version = "0.3.6"
@@ -3769,15 +3832,15 @@ dependencies = [
 
 [[package]]
 name = "serde_json"
-version = "1.0.145"
+version = "1.0.149"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "402a6f66d8c709116cf22f558eab210f5a50187f702eb4d7e5ef38d9a7f1c79c"
+checksum = "83fc039473c5595ace860d8c4fafa220ff474b3fc6bfdb4293327f1a37e94d86"
 dependencies = [
  "itoa",
  "memchr",
- "ryu",
  "serde",
  "serde_core",
+ "zmij",
 ]
 
 [[package]]
@@ -3829,24 +3892,25 @@ checksum = "0fda2ff0d084019ba4d7c6f371c95d8fd75ce3524c3cb8fb653a3023f6323e64"
 
 [[package]]
 name = "signal-hook-registry"
-version = "1.4.6"
+version = "1.4.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b2a4719bff48cee6b39d12c020eeb490953ad2443b7055bd0b21fca26bd8c28b"
+checksum = "c4db69cba1110affc0e9f7bcd48bbf87b3f4fc7c61fc9155afd4c469eb3d6c1b"
 dependencies = [
+ "errno",
  "libc",
 ]
 
 [[package]]
 name = "simd-adler32"
-version = "0.3.7"
+version = "0.3.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d66dc143e6b11c1eddc06d5c423cfc97062865baf299914ab64caa38182078fe"
+checksum = "703d5c7ef118737c72f1af64ad2f6f8c5e1921f818cdcb97b8fe6fc69bf66214"
 
 [[package]]
 name = "slab"
-version = "0.4.11"
+version = "0.4.12"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7a2ae44ef20feb57a68b23d846850f861394c2e02dc425a50098ae8c90267589"
+checksum = "0c790de23124f9ab44544d7ac05d60440adc586479ce501c1d6d7da3cd8c9cf5"
 
 [[package]]
 name = "smallvec"
@@ -3866,12 +3930,12 @@ dependencies = [
 
 [[package]]
 name = "socket2"
-version = "0.6.1"
+version = "0.6.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "17129e116933cf371d018bb80ae557e889637989d8638274fb25622827b03881"
+checksum = "3a766e1110788c36f4fa1c2b71b387a7815aa65f88ce0229841826633d93723e"
 dependencies = [
  "libc",
- "windows-sys 0.60.2",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
@@ -3966,9 +4030,9 @@ checksum = "13c2bddecc57b384dee18652358fb23172facb8a2c51ccc10d74c157bdea3292"
 
 [[package]]
 name = "syn"
-version = "2.0.107"
+version = "2.0.117"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2a26dbd934e5451d21ef060c018dae56fc073894c5a7896f882928a76e6d081b"
+checksum = "e665b8803e7b1d2a727f4023456bbbbe74da67099c585258af0ad9c5013b9b99"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -4007,7 +4071,7 @@ version = "0.6.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "01198a2debb237c62b6826ec7081082d951f46dbb64b0e8c7649a452230d1dfc"
 dependencies = [
- "bitflags 2.10.0",
+ "bitflags 2.11.1",
  "byteorder",
  "enum-as-inner",
  "libc",
@@ -4038,9 +4102,9 @@ dependencies = [
 
 [[package]]
 name = "tar"
-version = "0.4.44"
+version = "0.4.45"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1d863878d212c87a19c1a610eb53bb01fe12951c0501cf5a0d65f724914a667a"
+checksum = "22692a6476a21fa75fdfc11d452fda482af402c008cdbaf3476414e122040973"
 dependencies = [
  "filetime",
  "libc",
@@ -4049,15 +4113,15 @@ dependencies = [
 
 [[package]]
 name = "tempfile"
-version = "3.23.0"
+version = "3.27.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2d31c77bdf42a745371d260a26ca7163f1e0924b64afa0b688e61b5a9fa02f16"
+checksum = "32497e9a4c7b38532efcdebeef879707aa9f794296a4f0244f6f69e9bc8574bd"
 dependencies = [
  "fastrand",
- "getrandom 0.3.4",
+ "getrandom 0.4.2",
  "once_cell",
  "rustix",
- "windows-sys 0.60.2",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
@@ -4071,11 +4135,11 @@ dependencies = [
 
 [[package]]
 name = "thiserror"
-version = "2.0.17"
+version = "2.0.18"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f63587ca0f12b72a0600bcba1d40081f830876000bb46dd2337a3051618f4fc8"
+checksum = "4288b5bcbc7920c07a1149a35cf9590a2aa808e0bc1eafaade0b80947865fbc4"
 dependencies = [
- "thiserror-impl 2.0.17",
+ "thiserror-impl 2.0.18",
 ]
 
 [[package]]
@@ -4091,9 +4155,9 @@ dependencies = [
 
 [[package]]
 name = "thiserror-impl"
-version = "2.0.17"
+version = "2.0.18"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3ff15c8ecd7de3849db632e14d18d2571fa09dfc5ed93479bc4485c7a517c913"
+checksum = "ebc4ee7f67670e9b64d05fa4253e753e016c6c95ff35b89b7941d6b856dec1d5"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -4117,30 +4181,30 @@ dependencies = [
 
 [[package]]
 name = "time"
-version = "0.3.44"
+version = "0.3.47"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "91e7d9e3bb61134e77bde20dd4825b97c010155709965fedf0f49bb138e52a9d"
+checksum = "743bd48c283afc0388f9b8827b976905fb217ad9e647fae3a379a9283c4def2c"
 dependencies = [
  "deranged",
  "itoa",
  "num-conv",
  "powerfmt",
- "serde",
+ "serde_core",
  "time-core",
  "time-macros",
 ]
 
 [[package]]
 name = "time-core"
-version = "0.1.6"
+version = "0.1.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "40868e7c1d2f0b8d73e4a8c7f0ff63af4f6d19be117e90bd73eb1d62cf831c6b"
+checksum = "7694e1cfe791f8d31026952abf09c69ca6f6fa4e1a1229e18988f06a04a12dca"
 
 [[package]]
 name = "time-macros"
-version = "0.2.24"
+version = "0.2.27"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "30cfb0125f12d9c277f35663a0a33f8c30190f4e4574868a330595412d34ebf3"
+checksum = "2e70e4c5a0e0a8a4823ad65dfe1a6930e4f4d756dcd9dd7939022b5e8c501215"
 dependencies = [
  "num-conv",
  "time-core",
@@ -4148,9 +4212,9 @@ dependencies = [
 
 [[package]]
 name = "tinystr"
-version = "0.8.1"
+version = "0.8.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5d4f6d1145dcb577acf783d4e601bc1d76a13337bb54e6233add580b07344c8b"
+checksum = "c8323304221c2a851516f22236c5722a72eaa19749016521d6dff0824447d96d"
 dependencies = [
  "displaydoc",
  "zerovec",
@@ -4158,9 +4222,9 @@ dependencies = [
 
 [[package]]
 name = "tinyvec"
-version = "1.10.0"
+version = "1.11.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bfa5fdc3bce6191a1dbc8c02d5c8bffcf557bafa17c124c5264a458f1b0613fa"
+checksum = "3e61e67053d25a4e82c844e8424039d9745781b3fc4f32b8d55ed50f5f667ef3"
 dependencies = [
  "tinyvec_macros",
 ]
@@ -4173,9 +4237,9 @@ checksum = "1f3ccbac311fea05f86f61904b462b55fb3df8837a366dfc601a0161d0532f20"
 
 [[package]]
 name = "tokenizers"
-version = "0.22.1"
+version = "0.22.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6475a27088c98ea96d00b39a9ddfb63780d1ad4cceb6f48374349a96ab2b7842"
+checksum = "b238e22d44a15349529690fb07bd645cf58149a1b1e44d6cb5bd1641ff1a6223"
 dependencies = [
  "ahash",
  "aho-corasick",
@@ -4184,22 +4248,22 @@ dependencies = [
  "derive_builder",
  "esaxx-rs",
  "getrandom 0.3.4",
- "indicatif",
+ "indicatif 0.18.4",
  "itertools",
  "log",
  "macro_rules_attribute",
  "monostate",
  "onig",
  "paste",
- "rand 0.9.2",
+ "rand 0.9.4",
  "rayon",
  "rayon-cond",
  "regex",
- "regex-syntax 0.8.8",
+ "regex-syntax 0.8.10",
  "serde",
  "serde_json",
  "spm_precompiled",
- "thiserror 2.0.17",
+ "thiserror 2.0.18",
  "unicode-normalization-alignments",
  "unicode-segmentation",
  "unicode_categories",
@@ -4207,25 +4271,25 @@ dependencies = [
 
 [[package]]
 name = "tokio"
-version = "1.48.0"
+version = "1.52.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ff360e02eab121e0bc37a2d3b4d4dc622e6eda3a8e5253d5435ecf5bd4c68408"
+checksum = "8fc7f01b389ac15039e4dc9531aa973a135d7a4135281b12d7c1bc79fd57fffe"
 dependencies = [
  "bytes",
  "libc",
  "mio",
  "pin-project-lite",
  "signal-hook-registry",
- "socket2 0.6.1",
+ "socket2 0.6.3",
  "tokio-macros",
  "windows-sys 0.61.2",
 ]
 
 [[package]]
 name = "tokio-macros"
-version = "2.6.0"
+version = "2.7.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "af407857209536a95c8e56f8231ef2c2e2aff839b22e07a1ffcbc617e9db9fa5"
+checksum = "385a6cb71ab9ab790c5fe8d67f1645e6c450a7ce006a33de03daa956cf70a496"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -4248,15 +4312,15 @@ version = "0.26.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1729aa945f29d91ba541258c8df89027d5792d85a8841fb65e8bf0f4ede4ef61"
 dependencies = [
- "rustls 0.23.34",
+ "rustls 0.23.40",
  "tokio",
 ]
 
 [[package]]
 name = "tokio-stream"
-version = "0.1.17"
+version = "0.1.18"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "eca58d7bba4a75707817a2c44174253f9236b2d5fbd055602e9d5c07c139a047"
+checksum = "32da49809aab5c3bc678af03902d4ccddea2a87d028d86392a4b1560c6906c70"
 dependencies = [
  "futures-core",
  "pin-project-lite",
@@ -4265,9 +4329,9 @@ dependencies = [
 
 [[package]]
 name = "tokio-util"
-version = "0.7.16"
+version = "0.7.18"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "14307c986784f72ef81c89db7d9e28d6ac26d16213b109ea501696195e6e3ce5"
+checksum = "9ae9cec805b01e8fc3fd2fe289f89149a9b66dd16786abd8b19cfa7b48cb0098"
 dependencies = [
  "bytes",
  "futures-core",
@@ -4319,9 +4383,9 @@ checksum = "5d99f8c9a7727884afe522e9bd5edbfc91a3312b36a77b5fb8926e4c31a41801"
 
 [[package]]
 name = "tower"
-version = "0.5.2"
+version = "0.5.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d039ad9159c98b70ecfd540b2573b97f7f52c3e8d9f8ad57a24b916a536975f9"
+checksum = "ebe5ef63511595f1344e2d5cfa636d973292adc0eec1f0ad45fae9f0851ab1d4"
 dependencies = [
  "futures-core",
  "futures-util",
@@ -4334,20 +4398,20 @@ dependencies = [
 
 [[package]]
 name = "tower-http"
-version = "0.6.6"
+version = "0.6.10"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "adc82fd73de2a9722ac5da747f12383d2bfdb93591ee6c58486e0097890f05f2"
+checksum = "68d6fdd9f81c2819c9a8b0e0cd91660e7746a8e6ea2ba7c6b2b057985f6bcb51"
 dependencies = [
- "bitflags 2.10.0",
+ "bitflags 2.11.1",
  "bytes",
  "futures-util",
- "http 1.3.1",
+ "http 1.4.0",
  "http-body 1.0.1",
- "iri-string",
  "pin-project-lite",
  "tower",
  "tower-layer",
  "tower-service",
+ "url",
 ]
 
 [[package]]
@@ -4364,9 +4428,9 @@ checksum = "8df9b6e13f2d32c91b9bd719c00d1958837bc7dec474d94952798cc8e69eeec3"
 
 [[package]]
 name = "tracing"
-version = "0.1.41"
+version = "0.1.44"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "784e0ac535deb450455cbfa28a6f0df145ea1bb7ae51b821cf5e7927fdcfbdd0"
+checksum = "63e71662fa4b2a2c3a26f570f037eb95bb1f85397f3cd8076caed2f026a6d100"
 dependencies = [
  "pin-project-lite",
  "tracing-attributes",
@@ -4375,9 +4439,9 @@ dependencies = [
 
 [[package]]
 name = "tracing-attributes"
-version = "0.1.30"
+version = "0.1.31"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "81383ab64e72a7a8b8e13130c49e3dab29def6d0c7d76a03087b3cf71c5c6903"
+checksum = "7490cfa5ec963746568740651ac6781f701c9c5ea257c58e057f3ba8cf69e8da"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -4386,9 +4450,9 @@ dependencies = [
 
 [[package]]
 name = "tracing-core"
-version = "0.1.34"
+version = "0.1.36"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b9d12581f227e93f094d3af2ae690a574abb8a2b9b7a96e7cfe9647b2b617678"
+checksum = "db97caf9d906fbde555dd62fa95ddba9eecfd14cb388e4f491a66d74cd5fb79a"
 dependencies = [
  "once_cell",
  "valuable",
@@ -4431,15 +4495,15 @@ checksum = "e421abadd41a4225275504ea4d6566923418b7f05506fbc9c0fe86ba7396114b"
 
 [[package]]
 name = "typed-path"
-version = "0.12.2"
+version = "0.12.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3015e6ce46d5ad8751e4a772543a30c7511468070e98e64e20165f8f81155b64"
+checksum = "8e28f89b80c87b8fb0cf04ab448d5dd0dd0ade2f8891bae878de66a75a28600e"
 
 [[package]]
 name = "typenum"
-version = "1.19.0"
+version = "1.20.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "562d481066bde0658276a35467c4af00bdc6ee726305698a55b86e61d7ad82bb"
+checksum = "40ce102ab67701b8526c123c1bab5cbe42d7040ccfd0f64af1a385808d2f43de"
 
 [[package]]
 name = "ubyte"
@@ -4516,9 +4580,9 @@ dependencies = [
 
 [[package]]
 name = "unicode-ident"
-version = "1.0.19"
+version = "1.0.24"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f63a545481291138910575129486daeaf8ac54aee4387fe7906919f7830c7d9d"
+checksum = "e6e4313cd5fcd3dad5cafa179702e2b244f760991f45397d14d4ebf38247da75"
 
 [[package]]
 name = "unicode-normalization-alignments"
@@ -4531,9 +4595,9 @@ dependencies = [
 
 [[package]]
 name = "unicode-segmentation"
-version = "1.12.0"
+version = "1.13.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f6ccf251212114b54433ec949fd6a7841275f9ada20dddd2f29e9ceea4501493"
+checksum = "9629274872b2bfaf8d66f5f15725007f635594914870f65218920345aa11aa8c"
 
 [[package]]
 name = "unicode-width"
@@ -4553,6 +4617,12 @@ version = "0.1.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "39ec24b3121d976906ece63c9daad25b85969647682eee313cb5779fdd69e14e"
 
+[[package]]
+name = "unit-prefix"
+version = "0.5.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "81e544489bf3d8ef66c953931f56617f423cd4b5494be343d9b9d3dda037b9a3"
+
 [[package]]
 name = "untrusted"
 version = "0.9.0"
@@ -4569,7 +4639,7 @@ dependencies = [
  "flate2",
  "log",
  "once_cell",
- "rustls 0.23.34",
+ "rustls 0.23.40",
  "rustls-pki-types",
  "serde",
  "serde_json",
@@ -4580,9 +4650,9 @@ dependencies = [
 
 [[package]]
 name = "url"
-version = "2.5.7"
+version = "2.5.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "08bc136a29a3d1758e07a9cca267be308aeebf5cfd5a10f3f67ab2097683ef5b"
+checksum = "ff67a8a4397373c3ef660812acab3268222035010ab8680ec4215f38ba3d0eed"
 dependencies = [
  "form_urlencoded",
  "idna",
@@ -4604,11 +4674,11 @@ checksum = "06abde3611657adf66d383f00b093d7faecc7fa57071cce2578660c9f1010821"
 
 [[package]]
 name = "uuid"
-version = "1.18.1"
+version = "1.23.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2f87b8aa10b915a06587d0dec516c282ff295b475d94abf425d62b57710070a2"
+checksum = "ddd74a9687298c6858e9b88ec8935ec45d22e8fd5e6394fa1bd4e99a87789c76"
 dependencies = [
- "getrandom 0.3.4",
+ "getrandom 0.4.2",
  "js-sys",
  "wasm-bindgen",
 ]
@@ -4652,58 +4722,50 @@ checksum = "ccf3ec651a847eb01de73ccad15eb7d99f80485de043efb2f370cd654f4ea44b"
 
 [[package]]
 name = "wasip2"
-version = "1.0.1+wasi-0.2.4"
+version = "1.0.3+wasi-0.2.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0562428422c63773dad2c345a1882263bbf4d65cf3f42e90921f787ef5ad58e7"
+checksum = "20064672db26d7cdc89c7798c48a0fdfac8213434a1186e5ef29fd560ae223d6"
 dependencies = [
- "wit-bindgen",
+ "wit-bindgen 0.57.1",
 ]
 
 [[package]]
-name = "wasm-bindgen"
-version = "0.2.104"
+name = "wasip3"
+version = "0.4.0+wasi-0.3.0-rc-2026-01-06"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c1da10c01ae9f1ae40cbfac0bac3b1e724b320abfcf52229f80b547c0d250e2d"
+checksum = "5428f8bf88ea5ddc08faddef2ac4a67e390b88186c703ce6dbd955e1c145aca5"
 dependencies = [
- "cfg-if",
- "once_cell",
- "rustversion",
- "wasm-bindgen-macro",
- "wasm-bindgen-shared",
+ "wit-bindgen 0.51.0",
 ]
 
 [[package]]
-name = "wasm-bindgen-backend"
-version = "0.2.104"
+name = "wasm-bindgen"
+version = "0.2.121"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "671c9a5a66f49d8a47345ab942e2cb93c7d1d0339065d4f8139c486121b43b19"
+checksum = "49ace1d07c165b0864824eee619580c4689389afa9dc9ed3a4c75040d82e6790"
 dependencies = [
- "bumpalo",
- "log",
- "proc-macro2",
- "quote",
- "syn",
+ "cfg-if",
+ "once_cell",
+ "rustversion",
+ "wasm-bindgen-macro",
  "wasm-bindgen-shared",
 ]
 
 [[package]]
 name = "wasm-bindgen-futures"
-version = "0.4.54"
+version = "0.4.71"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7e038d41e478cc73bae0ff9b36c60cff1c98b8f38f8d7e8061e79ee63608ac5c"
+checksum = "96492d0d3ffba25305a7dc88720d250b1401d7edca02cc3bcd50633b424673b8"
 dependencies = [
- "cfg-if",
  "js-sys",
- "once_cell",
  "wasm-bindgen",
- "web-sys",
 ]
 
 [[package]]
 name = "wasm-bindgen-macro"
-version = "0.2.104"
+version = "0.2.121"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7ca60477e4c59f5f2986c50191cd972e3a50d8a95603bc9434501cf156a9a119"
+checksum = "8e68e6f4afd367a562002c05637acb8578ff2dea1943df76afb9e83d177c8578"
 dependencies = [
  "quote",
  "wasm-bindgen-macro-support",
@@ -4711,26 +4773,48 @@ dependencies = [
 
 [[package]]
 name = "wasm-bindgen-macro-support"
-version = "0.2.104"
+version = "0.2.121"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9f07d2f20d4da7b26400c9f4a0511e6e0345b040694e8a75bd41d578fa4421d7"
+checksum = "d95a9ec35c64b2a7cb35d3fead40c4238d0940c86d107136999567a4703259f2"
 dependencies = [
+ "bumpalo",
  "proc-macro2",
  "quote",
  "syn",
- "wasm-bindgen-backend",
  "wasm-bindgen-shared",
 ]
 
 [[package]]
 name = "wasm-bindgen-shared"
-version = "0.2.104"
+version = "0.2.121"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bad67dc8b2a1a6e5448428adec4c3e84c43e561d8c9ee8a9e5aabeb193ec41d1"
+checksum = "c4e0100b01e9f0d03189a92b96772a1fb998639d981193d7dbab487302513441"
 dependencies = [
  "unicode-ident",
 ]
 
+[[package]]
+name = "wasm-encoder"
+version = "0.244.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "990065f2fe63003fe337b932cfb5e3b80e0b4d0f5ff650e6985b1048f62c8319"
+dependencies = [
+ "leb128fmt",
+ "wasmparser",
+]
+
+[[package]]
+name = "wasm-metadata"
+version = "0.244.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bb0e353e6a2fbdc176932bbaab493762eb1255a7900fe0fea1a2f96c296cc909"
+dependencies = [
+ "anyhow",
+ "indexmap",
+ "wasm-encoder",
+ "wasmparser",
+]
+
 [[package]]
 name = "wasm-streams"
 version = "0.4.2"
@@ -4744,11 +4828,23 @@ dependencies = [
  "web-sys",
 ]
 
+[[package]]
+name = "wasmparser"
+version = "0.244.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "47b807c72e1bac69382b3a6fb3dbe8ea4c0ed87ff5629b8685ae6b9a611028fe"
+dependencies = [
+ "bitflags 2.11.1",
+ "hashbrown 0.15.5",
+ "indexmap",
+ "semver",
+]
+
 [[package]]
 name = "web-sys"
-version = "0.3.81"
+version = "0.3.98"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9367c417a924a74cae129e6a2ae3b47fabb1f8995595ab474029da749a8be120"
+checksum = "4b572dff8bcf38bad0fa19729c89bb5748b2b9b1d8be70cf90df697e3a8f32aa"
 dependencies = [
  "js-sys",
  "wasm-bindgen",
@@ -4776,14 +4872,14 @@ version = "0.26.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "521bc38abb08001b01866da9f51eb7c5d647a19260e00054a8c7fd5f9e57f7a9"
 dependencies = [
- "webpki-roots 1.0.3",
+ "webpki-roots 1.0.7",
 ]
 
 [[package]]
 name = "webpki-roots"
-version = "1.0.3"
+version = "1.0.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "32b130c0d2d49f8b6889abc456e795e82525204f27c42cf767cf0d7734e089b8"
+checksum = "52f5ee44c96cf55f1b349600768e3ece3a8f26010c05265ab73f945bb1a2eb9d"
 dependencies = [
  "rustls-pki-types",
 ]
@@ -4810,7 +4906,7 @@ version = "0.1.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "c2a7b1c03c876122aa43f3020e6c3c3ee5c05081c9a00739faf7503aeba10d22"
 dependencies = [
- "windows-sys 0.60.2",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
@@ -4830,22 +4926,22 @@ dependencies = [
 
 [[package]]
 name = "windows-core"
-version = "0.61.2"
+version = "0.62.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c0fdd3ddb90610c7638aa2b3a3ab2904fb9e5cdbecc643ddb3647212781c4ae3"
+checksum = "b8e83a14d34d0623b51dce9581199302a221863196a1dde71a7663a4c2be9deb"
 dependencies = [
  "windows-implement",
  "windows-interface",
- "windows-link 0.1.3",
+ "windows-link",
  "windows-result",
  "windows-strings",
 ]
 
 [[package]]
 name = "windows-implement"
-version = "0.60.0"
+version = "0.60.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a47fddd13af08290e67f4acabf4b459f647552718f683a7b415d290ac744a836"
+checksum = "053e2e040ab57b9dc951b72c264860db7eb3b0200ba345b4e4c3b14f67855ddf"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -4854,21 +4950,15 @@ dependencies = [
 
 [[package]]
 name = "windows-interface"
-version = "0.59.1"
+version = "0.59.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bd9211b69f8dcdfa817bfd14bf1c97c9188afa36f4750130fcdf3f400eca9fa8"
+checksum = "3f316c4a2570ba26bbec722032c4099d8c8bc095efccdc15688708623367e358"
 dependencies = [
  "proc-macro2",
  "quote",
  "syn",
 ]
 
-[[package]]
-name = "windows-link"
-version = "0.1.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5e6ad25900d524eaabdbbb96d20b4311e1e7ae1699af4fb28c17ae66c80d798a"
-
 [[package]]
 name = "windows-link"
 version = "0.2.1"
@@ -4877,20 +4967,20 @@ checksum = "f0805222e57f7521d6a62e36fa9163bc891acd422f971defe97d64e70d0a4fe5"
 
 [[package]]
 name = "windows-result"
-version = "0.3.4"
+version = "0.4.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "56f42bd332cc6c8eac5af113fc0c1fd6a8fd2aa08a0119358686e5160d0586c6"
+checksum = "7781fa89eaf60850ac3d2da7af8e5242a5ea78d1a11c49bf2910bb5a73853eb5"
 dependencies = [
- "windows-link 0.1.3",
+ "windows-link",
 ]
 
 [[package]]
 name = "windows-strings"
-version = "0.4.2"
+version = "0.5.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "56e6c93f3a0c3b36176cb1327a4958a0353d5d166c2a35cb268ace15e91d3b57"
+checksum = "7837d08f69c77cf6b07689544538e017c1bfcf57e34b4c0ff58e6c2cd3b37091"
 dependencies = [
- "windows-link 0.1.3",
+ "windows-link",
 ]
 
 [[package]]
@@ -4935,7 +5025,7 @@ version = "0.61.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ae137229bcbd6cdf0f7b80a31df61766145077ddf49416a728b02cb3921ff3fc"
 dependencies = [
- "windows-link 0.2.1",
+ "windows-link",
 ]
 
 [[package]]
@@ -4975,7 +5065,7 @@ version = "0.53.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "4945f9f551b88e0d65f3db0bc25c33b8acea4d9e41163edf90dcd0b19f9069f3"
 dependencies = [
- "windows-link 0.2.1",
+ "windows-link",
  "windows_aarch64_gnullvm 0.53.1",
  "windows_aarch64_msvc 0.53.1",
  "windows_i686_gnu 0.53.1",
@@ -5126,9 +5216,9 @@ checksum = "d6bbff5f0aada427a1e5a6da5f1f98158182f26556f345ac9e04d36d0ebed650"
 
 [[package]]
 name = "winnow"
-version = "0.7.13"
+version = "0.7.15"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "21a0236b59786fed61e2a80582dd500fe61f18b5dca67a4a067d0bc9039339cf"
+checksum = "df79d97927682d2fd8adb29682d1140b343be4ac0f08fd68b7765d9c059d3945"
 dependencies = [
  "memchr",
 ]
@@ -5145,21 +5235,109 @@ dependencies = [
 
 [[package]]
 name = "wit-bindgen"
-version = "0.46.0"
+version = "0.51.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d7249219f66ced02969388cf2bb044a09756a083d0fab1e566056b04d9fbcaa5"
+dependencies = [
+ "wit-bindgen-rust-macro",
+]
+
+[[package]]
+name = "wit-bindgen"
+version = "0.57.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1ebf944e87a7c253233ad6766e082e3cd714b5d03812acc24c318f549614536e"
+
+[[package]]
+name = "wit-bindgen-core"
+version = "0.51.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ea61de684c3ea68cb082b7a88508a8b27fcc8b797d738bfc99a82facf1d752dc"
+dependencies = [
+ "anyhow",
+ "heck",
+ "wit-parser",
+]
+
+[[package]]
+name = "wit-bindgen-rust"
+version = "0.51.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b7c566e0f4b284dd6561c786d9cb0142da491f46a9fbed79ea69cdad5db17f21"
+dependencies = [
+ "anyhow",
+ "heck",
+ "indexmap",
+ "prettyplease",
+ "syn",
+ "wasm-metadata",
+ "wit-bindgen-core",
+ "wit-component",
+]
+
+[[package]]
+name = "wit-bindgen-rust-macro"
+version = "0.51.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0c0f9bfd77e6a48eccf51359e3ae77140a7f50b1e2ebfe62422d8afdaffab17a"
+dependencies = [
+ "anyhow",
+ "prettyplease",
+ "proc-macro2",
+ "quote",
+ "syn",
+ "wit-bindgen-core",
+ "wit-bindgen-rust",
+]
+
+[[package]]
+name = "wit-component"
+version = "0.244.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f17a85883d4e6d00e8a97c586de764dabcc06133f7f1d55dce5cdc070ad7fe59"
+checksum = "9d66ea20e9553b30172b5e831994e35fbde2d165325bec84fc43dbf6f4eb9cb2"
+dependencies = [
+ "anyhow",
+ "bitflags 2.11.1",
+ "indexmap",
+ "log",
+ "serde",
+ "serde_derive",
+ "serde_json",
+ "wasm-encoder",
+ "wasm-metadata",
+ "wasmparser",
+ "wit-parser",
+]
+
+[[package]]
+name = "wit-parser"
+version = "0.244.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ecc8ac4bc1dc3381b7f59c34f00b67e18f910c2c0f50015669dde7def656a736"
+dependencies = [
+ "anyhow",
+ "id-arena",
+ "indexmap",
+ "log",
+ "semver",
+ "serde",
+ "serde_derive",
+ "serde_json",
+ "unicode-xid",
+ "wasmparser",
+]
 
 [[package]]
 name = "writeable"
-version = "0.6.1"
+version = "0.6.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ea2f10b9bb0928dfb1b42b65e1f9e36f7f54dbdf08457afefb38afcdec4fa2bb"
+checksum = "1ffae5123b2d3fc086436f8834ae3ab053a283cfac8fe0a0b8eaae044768a4c4"
 
 [[package]]
 name = "xattr"
-version = "1.5.0"
+version = "1.6.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0d65cbf2f12c15564212d48f4e3dfb87923d25d611f2aed18f4cb23f0413d89e"
+checksum = "32e45ad4206f6d2479085147f02bc2ef834ac85886624a23575ae137c8aa8156"
 dependencies = [
  "libc",
  "rustix",
@@ -5188,12 +5366,12 @@ dependencies = [
 
 [[package]]
 name = "yoke"
-version = "0.8.1"
+version = "0.8.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "72d6e5c6afb84d73944e5cedb052c4680d5657337201555f9f2a16b7406d4954"
+checksum = "abe8c5fda708d9ca3df187cae8bfb9ceda00dd96231bed36e445a1a48e66f9ca"
 dependencies = [
  "stable_deref_trait",
- "yoke-derive 0.8.0",
+ "yoke-derive 0.8.2",
  "zerofrom",
 ]
 
@@ -5211,9 +5389,9 @@ dependencies = [
 
 [[package]]
 name = "yoke-derive"
-version = "0.8.0"
+version = "0.8.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "38da3c9736e16c5d3c8c597a9aaa5d1fa565d0532ae05e27c24aa62fb32c0ab6"
+checksum = "de844c262c8848816172cef550288e7dc6c7b7814b4ee56b3e1553f275f1858e"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -5223,18 +5401,18 @@ dependencies = [
 
 [[package]]
 name = "zerocopy"
-version = "0.8.27"
+version = "0.8.48"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0894878a5fa3edfd6da3f88c4805f4c8558e2b996227a3d864f47fe11e38282c"
+checksum = "eed437bf9d6692032087e337407a86f04cd8d6a16a37199ed57949d415bd68e9"
 dependencies = [
  "zerocopy-derive",
 ]
 
 [[package]]
 name = "zerocopy-derive"
-version = "0.8.27"
+version = "0.8.48"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "88d2b8d9c68ad2b9e4340d7832716a4d21a22a1154777ad56ea55c51a9cf3831"
+checksum = "70e3cd084b1788766f53af483dd21f93881ff30d7320490ec3ef7526d203bad4"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -5243,18 +5421,18 @@ dependencies = [
 
 [[package]]
 name = "zerofrom"
-version = "0.1.6"
+version = "0.1.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "50cc42e0333e05660c3587f3bf9d0478688e15d870fab3346451ce7f8c9fbea5"
+checksum = "0ec05a11813ea801ff6d75110ad09cd0824ddba17dfe17128ea0d5f68e6c5272"
 dependencies = [
  "zerofrom-derive",
 ]
 
 [[package]]
 name = "zerofrom-derive"
-version = "0.1.6"
+version = "0.1.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d71e5d6e06ab090c67b5e44993ec16b72dcbaabc526db883a360057678b48502"
+checksum = "11532158c46691caf0f2593ea8358fed6bbf68a0315e80aae9bd41fbade684a1"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -5270,31 +5448,31 @@ checksum = "b97154e67e32c85465826e8bcc1c59429aaaf107c1e4a9e53c8d8ccd5eff88d0"
 
 [[package]]
 name = "zerotrie"
-version = "0.2.2"
+version = "0.2.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "36f0bbd478583f79edad978b407914f61b2972f5af6fa089686016be8f9af595"
+checksum = "0f9152d31db0792fa83f70fb2f83148effb5c1f5b8c7686c3459e361d9bc20bf"
 dependencies = [
  "displaydoc",
- "yoke 0.8.1",
+ "yoke 0.8.2",
  "zerofrom",
 ]
 
 [[package]]
 name = "zerovec"
-version = "0.11.4"
+version = "0.11.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e7aa2bd55086f1ab526693ecbe444205da57e25f4489879da80635a46d90e73b"
+checksum = "90f911cbc359ab6af17377d242225f4d75119aec87ea711a880987b18cd7b239"
 dependencies = [
- "yoke 0.8.1",
+ "yoke 0.8.2",
  "zerofrom",
  "zerovec-derive",
 ]
 
 [[package]]
 name = "zerovec-derive"
-version = "0.11.1"
+version = "0.11.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5b96237efa0c878c64bd89c436f661be4e46b2f3eff1ebb976f7ef2321d2f58f"
+checksum = "625dc425cab0dca6dc3c3319506e6593dcb08a9f387ea3b284dbd52a92c40555"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -5314,15 +5492,15 @@ dependencies = [
  "flate2",
  "indexmap",
  "memchr",
- "thiserror 2.0.17",
+ "thiserror 2.0.18",
  "zopfli",
 ]
 
 [[package]]
 name = "zip"
-version = "7.4.0"
+version = "7.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cc12baa6db2b15a140161ce53d72209dacea594230798c24774139b54ecaa980"
+checksum = "c42e33efc22a0650c311c2ef19115ce232583abbe80850bc8b66509ebef02de0"
 dependencies = [
  "crc32fast",
  "indexmap",
@@ -5330,11 +5508,17 @@ dependencies = [
  "typed-path",
 ]
 
+[[package]]
+name = "zmij"
+version = "1.0.21"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b8848ee67ecc8aedbaf3e4122217aff892639231befc6a1b58d29fff4c2cabaa"
+
 [[package]]
 name = "zopfli"
-version = "0.8.2"
+version = "0.8.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "edfc5ee405f504cd4984ecc6f14d02d55cfda60fa4b689434ef4102aae150cd7"
+checksum = "f05cd8797d63865425ff89b5c4a48804f35ba0ce8d125800027ad6017d2b5249"
 dependencies = [
  "bumpalo",
  "crc32fast",
@@ -5344,15 +5528,35 @@ dependencies = [
 
 [[package]]
 name = "zune-core"
-version = "0.4.12"
+version = "0.5.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3f423a2c17029964870cfaabb1f13dfab7d092a62a29a89264f4d36990ca414a"
+checksum = "cb8a0807f7c01457d0379ba880ba6322660448ddebc890ce29bb64da71fb40f9"
 
 [[package]]
 name = "zune-jpeg"
-version = "0.4.21"
+version = "0.5.15"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "29ce2c8a9384ad323cf564b67da86e21d3cfdff87908bc1223ed5c99bc792713"
+checksum = "27bc9d5b815bc103f142aa054f561d9187d191692ec7c2d1e2b4737f8dbd7296"
 dependencies = [
  "zune-core",
 ]
+
+[[patch.unused]]
+name = "candle-core"
+version = "0.10.2"
+source = "git+https://github.com/huggingface/candle?branch=main#3df8203a2ab0f7d12866ef392d5ea7504b0255e4"
+
+[[patch.unused]]
+name = "candle-flash-attn"
+version = "0.10.2"
+source = "git+https://github.com/huggingface/candle?branch=main#3df8203a2ab0f7d12866ef392d5ea7504b0255e4"
+
+[[patch.unused]]
+name = "candle-nn"
+version = "0.10.2"
+source = "git+https://github.com/huggingface/candle?branch=main#3df8203a2ab0f7d12866ef392d5ea7504b0255e4"
+
+[[patch.unused]]
+name = "candle-transformers"
+version = "0.10.2"
+source = "git+https://github.com/huggingface/candle?branch=main#3df8203a2ab0f7d12866ef392d5ea7504b0255e4"
diff --git a/Cargo.toml b/Cargo.toml
index 0d25a4c..3feec4d 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -44,3 +44,11 @@ fast_image_resize = "5.3"
 # opt-level = 3
 # lto = true
 # codegen-units = 1
+[patch.crates-io]
+candle-core = { git = "https://github.com/huggingface/candle", branch = "main" }
+candle-nn = { git = "https://github.com/huggingface/candle", branch = "main" }
+candle-transformers = { git = "https://github.com/huggingface/candle", branch = "main" }
+candle-flash-attn = { git = "https://github.com/huggingface/candle", branch = "main" }
+
+# Force cudarc to a version that supports CUDA 13
+cudarc = { git = "https://github.com/coreylowman/cudarc", branch = "main" }
diff --git a/README.md b/README.md
index 5fedd13..e8fb724 100644
--- a/README.md
+++ b/README.md
@@ -50,7 +50,7 @@ The original DeepSeek-OCR ships as a Python + Transformers stack—powerful, but
 - **One repo, two entrypoints** – a batteries-included CLI for batch jobs and a Rocket-based server that speaks `/v1/responses` and `/v1/chat/completions`.
 - **Works out of the box** – pulls model weights, configs, and tokenizer from whichever of Hugging Face or ModelScope responds fastest on first run.
 - **Optimised for Apple Silicon** – optional Metal backend with FP16 execution for real-time OCR on laptops.
-- **CUDA (alpha)** – experimental support via `--features cuda` + `--device cuda --dtype f16`; expect rough edges while we finish kernel coverage.
+- **CUDA** – NVIDIA GPU acceleration via `--features cuda` + `--device cuda --dtype f16`. VRAM swap (`--vision-offload sequential`) brings DeepSeek‑OCR Q4K down to **~2.3GB VRAM peak** on low-end GPUs (RTX 3050 4GB, GTX 1060, etc.). Full GPU vision mode (`--vision-offload full-gpu`) available for higher-VRAM cards. See the [VRAM swap section](#vram-swap--low-vram-gpus) for details.
 - **Intel MKL (preview)** – faster BLAS on x86 via `--features mkl` (install Intel oneMKL beforehand).
 - **OpenAI client compatibility** – drop-in replacement for popular SDKs; the server automatically collapses chat history to the latest user turn for OCR-friendly prompts.
 
@@ -81,7 +81,7 @@ The workspace exposes three base model IDs plus DSQ-quantized variants for DeepS
 - Rust 1.78+ (edition 2024 support)
 - Git
 - Optional: Apple Silicon running macOS 13+ for Metal acceleration
-- Optional: CUDA 12.2+ toolkit + driver for experimental NVIDIA GPU acceleration on Linux/Windows
+- Optional: CUDA 12.2+ toolkit + driver for NVIDIA GPU acceleration on Linux/Windows
 - Optional: Intel oneAPI MKL for preview x86 acceleration (see below)
 - (Recommended) Hugging Face account with `HF_TOKEN` when pulling from the `deepseek-ai/DeepSeek-OCR` repo (ModelScope is used automatically when it’s faster/reachable).
 
@@ -132,6 +132,9 @@ template = "plain"
 base_size = 1024
 image_size = 640
 crop_mode = true
+vision_swap = true
+patches_per_batch = 2
+vision_offload = "auto"
 max_new_tokens = 512
 use_cache = true
 
@@ -141,7 +144,7 @@ port = 8000
 ```
 
 - `[models]` picks the active model and lets you add more entries (each entry can point to its own config/tokenizer/weights).
-- `[inference]` controls notebook-friendly defaults shared by the CLI and server (device, template, vision sizing, decoding budget, cache usage).
+- `[inference]` controls notebook-friendly defaults shared by the CLI and server (device, template, vision sizing, decoding budget, cache usage, VRAM swap settings).
 - `[server]` sets the network binding and the model identifier reported by `/v1/models`.
 
 See `crates/cli/README.md` and `crates/server/README.md` for concise override tables.
@@ -175,7 +178,7 @@ cargo run -p deepseek-ocr-cli --release -- \
 
 > macOS tip: append `--features metal` to the `cargo run`/`cargo build` commands to compile with Accelerate + Metal backends.
 >
-> CUDA tip (Linux/Windows): append `--features cuda` and run with `--device cuda --dtype f16` to target NVIDIA GPUs—feature is still alpha, so be ready for quirks.
+> CUDA tip (Linux/Windows): append `--features cuda` and run with `--device cuda --dtype f16` to target NVIDIA GPUs. Add `--vision-offload sequential` for low-VRAM cards.
 >
 > Intel MKL preview: install Intel oneMKL, then build with `--features mkl` for faster CPU matmuls on x86.
 
@@ -195,6 +198,10 @@ Key flags:
 - Sampling controls: `--do-sample`, `--temperature`, `--top-p`, `--top-k`, `--repetition-penalty`, `--no-repeat-ngram-size`, `--seed`
   - By default decoding stays deterministic (`do_sample=false`, `temperature=0.0`, `no_repeat_ngram_size=20`)
   - To use stochastic sampling set `--do-sample true --temperature 0.8` (and optionally adjust the other knobs)
+- VRAM swap flags (CUDA only, DeepSeek‑OCR):
+  - `--vision-offload <STRATEGY>` – choose vision offloading mode: `auto` (default), `sequential`, `full-gpu`, `cpu`
+  - `--vision-swap <BOOL>` – enable/disable VRAM-aware loading (overridden by `--vision-offload`)
+  - `--patches-per-batch <N>` – patch batch size for CUDA processing (default: 2; smaller = less VRAM)
 
 ### Switching Models
 
@@ -241,11 +248,45 @@ Notes:
 ## GPU Acceleration ⚡
 
 - **Metal (macOS 13+ Apple Silicon)** – pass `--device metal --dtype f16` and build binaries with `--features metal` so Candle links against Accelerate + Metal.
-- **CUDA (alpha, NVIDIA GPUs)** – install CUDA 12.2+ toolkits, build with `--features cuda`, and launch the CLI/server with `--device cuda --dtype f16`; still experimental.
+- **CUDA (NVIDIA GPUs)** – install CUDA 12.2+ toolkits, build with `--features cuda`, and launch the CLI/server with `--device cuda --dtype f16`. Includes VRAM swap support for low-memory GPUs (RTX 3050 4GB, GTX 1060, etc.) and full-GPU vision mode for larger cards. See [VRAM Swap](#vram-swap--low-vram-gpus) below.
 - **Intel MKL (preview)** – install Intel oneMKL and build with `--features mkl` to speed up CPU workloads on x86.
 - For either backend, prefer release builds (e.g. `cargo build --release -p deepseek-ocr-cli --features metal|cuda`) to maximise throughput.
 - Combine GPU runs with `--max-new-tokens` and crop tuning flags to balance latency vs. quality.
 
+## VRAM Swap & Low-VRAM GPUs 🔄
+
+DeepSeek‑OCR with Q4K quantisation (`deepseek-ocr-q4k`) fits in **~950 MB** on CUDA, but the SAM+CLIP vision models require an additional **~2.1 GB**. On low-VRAM GPUs (≤6 GB), a **VRAM swap** mode loads one vision model at a time:
+
+| Mode | `--vision-offload` | Global View | Patch Crops | Peak VRAM (Q4K) |
+|------|-------------------|-------------|-------------|-----------------|
+| **Auto** (default) | `auto` | Auto-detects: uses Sequential on <6 GB | | ~2.3 GB |
+| **Sequential** | `sequential` | CPU (always) | CUDA (chunked) | ~2.3 GB |
+| **Full GPU** | `full-gpu` | CUDA | CUDA (chunked) | ~3.2 GB |
+| **CPU only** | `cpu` | CPU | CPU | ~1 GB |
+
+**Sequential mode** runs the global view on CPU while patch crops process on CUDA in parallel (scoped threads), giving the best VRAM/speed trade-off on tight GPUs.
+
+**Full GPU mode** keeps all vision on CUDA: loads SAM → processes all images → drops SAM → loads CLIP → processes with cached SAM outputs. Offers faster vision prefill when VRAM allows (~3.2 GB peak).
+
+Fine-tune VRAM vs. throughput with `--patches-per-batch`:
+```bash
+# Minimum VRAM (processes one patch at a time on CUDA)
+cargo run --release --features cuda -- --device cuda --dtype f16 --model deepseek-ocr-q4k \
+  --vision-offload sequential --patches-per-batch 1 --prompt "<image> OCR" --image doc.png
+
+# Faster (batches 4 patches per CUDA forward)
+cargo run --release --features cuda -- --device cuda --dtype f16 --model deepseek-ocr-q4k \
+  --vision-offload sequential --patches-per-batch 4 --prompt "<image> OCR" --image doc.png
+```
+
+Disable swap entirely (force CPU vision) when you have enough RAM:
+```bash
+cargo run --release --features cuda -- --device cuda --dtype f16 --model deepseek-ocr-q4k \
+  --vision-offload cpu --prompt "<image> OCR" --image doc.png
+```
+
+> **Note:** VRAM swap is only relevant for the DeepSeek‑OCR vision pipeline (SAM+CLIP). PaddleOCR‑VL and DotsOCR do not benefit since they use lighter or different vision towers.
+
 ## Repository Layout 🗂️
 
 - `crates/core` – shared inference pipeline, model loaders, conversation templates.
@@ -265,7 +306,8 @@ Detailed CLI usage lives in [`crates/cli/README.md`](crates/cli/README.md). The
 ## Roadmap 🗺️
 
 - ✅ Apple Metal backend with FP16 support and CLI/server parity on macOS.
-- ✅ NVIDIA CUDA backend (alpha) – build with `--features cuda`, run with `--device cuda --dtype f16` for Linux/Windows GPUs; polishing in progress.
+- ✅ NVIDIA CUDA backend – build with `--features cuda`, run with `--device cuda --dtype f16` for Linux/Windows GPUs. VRAM swap brings DeepSeek‑OCR Q4K down to ~2.3 GB on RTX 3050 4 GB.
+- ✅ **VRAM Swap** – sequential vision offload for low-VRAM CUDA GPUs via `--vision-offload sequential`, full-GPU mode via `--vision-offload full-gpu`, CPU-only mode via `--vision-offload cpu`.
 - 🔄 **Parity polish** – finish projector normalisation + crop tiling alignment; extend intermediate-tensor diff suite beyond the current sample baseline.
 - 🔄 **Grounding & streaming** – port the Python post-processing helpers (box extraction, markdown polish) and refine SSE streaming ergonomics.
 - 🔄 **Cross-platform acceleration** – continue tuning CUDA kernels, add automatic device detection across CPU/Metal/CUDA, and publish opt-in GPU benchmarks.
diff --git a/crates/assets/src/lib.rs b/crates/assets/src/lib.rs
index d80d2cf..a3dd55e 100644
--- a/crates/assets/src/lib.rs
+++ b/crates/assets/src/lib.rs
@@ -99,6 +99,26 @@ pub const MODEL_ASSETS: &[ModelAsset] = &[
 ];
 
 pub const QUANTIZED_MODEL_ASSETS: &[QuantizedModelAsset] = &[
+    QuantizedModelAsset {
+        id: "deepseek-ocr-q2k",
+        kind: ModelKind::Deepseek,
+        baseline_id: "deepseek-ocr",
+        snapshot: SnapshotAsset {
+            dtype: "Q2_K",
+            repo_id: "TimmyOVO/deepseek-ocr.rs",
+            filename: "DeepSeek-OCR.Q2_K.dsq",
+        },
+    },
+    QuantizedModelAsset {
+        id: "deepseek-ocr-q3k",
+        kind: ModelKind::Deepseek,
+        baseline_id: "deepseek-ocr",
+        snapshot: SnapshotAsset {
+            dtype: "Q3_K",
+            repo_id: "TimmyOVO/deepseek-ocr.rs",
+            filename: "DeepSeek-OCR.Q3_K.dsq",
+        },
+    },
     QuantizedModelAsset {
         id: "deepseek-ocr-q4k",
         kind: ModelKind::Deepseek,
diff --git a/crates/cli/src/args.rs b/crates/cli/src/args.rs
index a9bbb3f..7f19601 100644
--- a/crates/cli/src/args.rs
+++ b/crates/cli/src/args.rs
@@ -49,7 +49,7 @@ pub struct SnapshotArgs {
         long,
         value_name = "DTYPE",
         default_value = "Q8_0",
-        value_parser = ["Q8_0", "Q4_K", "Q6_K"],
+        value_parser = ["Q8_0", "Q2_K", "Q3_K", "Q4_K", "Q6_K"],
         help_heading = "Snapshot"
     )]
     pub dtype: String,
diff --git a/crates/config/src/args.rs b/crates/config/src/args.rs
index 05a33d7..1b9c4e7 100644
--- a/crates/config/src/args.rs
+++ b/crates/config/src/args.rs
@@ -1,7 +1,7 @@
 use std::path::PathBuf;
 
 use clap::Args;
-use deepseek_ocr_core::runtime::{DeviceKind, Precision};
+use deepseek_ocr_core::runtime::{DeviceKind, Precision, VisionOffload};
 
 use crate::config::{ConfigOverrides, ServerOverride};
 
@@ -54,6 +54,22 @@ pub struct CommonInferenceArgs {
     #[arg(long, help_heading = "Inference")]
     pub crop_mode: Option<bool>,
 
+    /// Enable VRAM swap for vision models (auto-detect by default).
+    #[arg(long, help_heading = "Inference")]
+    pub vision_swap: Option<bool>,
+
+    /// Patch batch size for VRAM swap (default: 2).
+    #[arg(long, help_heading = "Inference")]
+    pub patches_per_batch: Option<usize>,
+
+    /// Number of patches to process on CPU in sequential mode (default: 0 = all on GPU).
+    #[arg(long, help_heading = "Inference")]
+    pub cpu_patches: Option<usize>,
+
+    /// Vision offload strategy (auto, sequential, full-gpu, cpu).
+    #[arg(long, help_heading = "Inference")]
+    pub vision_offload: Option<VisionOffload>,
+
     /// Default max tokens budget.
     #[arg(long, help_heading = "Inference")]
     pub max_new_tokens: Option<usize>,
@@ -111,6 +127,10 @@ impl From<&CommonInferenceArgs> for crate::InferenceOverride {
             base_size: value.base_size,
             image_size: value.image_size,
             crop_mode: value.crop_mode,
+            vision_swap: value.vision_swap,
+            patches_per_batch: value.patches_per_batch,
+            cpu_patches: value.cpu_patches,
+            vision_offload: value.vision_offload,
             decode: deepseek_ocr_core::DecodeParametersPatch {
                 max_new_tokens: value.max_new_tokens,
                 do_sample: value.do_sample,
diff --git a/crates/config/src/config.rs b/crates/config/src/config.rs
index b9418ca..830cad6 100644
--- a/crates/config/src/config.rs
+++ b/crates/config/src/config.rs
@@ -7,7 +7,7 @@ use std::{
 use anyhow::{Context, Result, anyhow};
 use deepseek_ocr_core::{
     DecodeParameters, DecodeParametersPatch, ModelKind,
-    runtime::{DeviceKind, Precision},
+    runtime::{DeviceKind, Precision, VisionOffload},
 };
 use serde::{Deserialize, Serialize};
 
@@ -62,6 +62,12 @@ fn ensure_default_model_entries(entries: &mut BTreeMap<String, ModelEntry>) {
     entries
         .entry("glm-ocr".to_string())
         .or_insert_with(glm_ocr_entry);
+    entries
+        .entry("deepseek-ocr-q2k".to_string())
+        .or_insert_with(|| quantized_entry(ModelKind::Deepseek, "Q2_K", "deepseek-ocr"));
+    entries
+        .entry("deepseek-ocr-q3k".to_string())
+        .or_insert_with(|| quantized_entry(ModelKind::Deepseek, "Q3_K", "deepseek-ocr"));
     entries
         .entry("deepseek-ocr-q4k".to_string())
         .or_insert_with(|| quantized_entry(ModelKind::Deepseek, "Q4_K", "deepseek-ocr"));
@@ -139,7 +145,13 @@ fn ensure_model_defaults(entries: &mut BTreeMap<String, ModelEntry>) {
         fill_missing_model_defaults(entry, &ocr2_defaults);
     }
 
-    let quantized_deepseek_ids = ["deepseek-ocr-q4k", "deepseek-ocr-q6k", "deepseek-ocr-q8k"];
+    let quantized_deepseek_ids = [
+        "deepseek-ocr-q2k",
+        "deepseek-ocr-q3k",
+        "deepseek-ocr-q4k",
+        "deepseek-ocr-q6k",
+        "deepseek-ocr-q8k",
+    ];
     for model_id in quantized_deepseek_ids {
         if let Some(entry) = entries.get_mut(model_id) {
             fill_missing_model_defaults(entry, &ocr1_defaults);
@@ -202,6 +214,10 @@ pub struct InferenceSettings {
     pub base_size: u32,
     pub image_size: u32,
     pub crop_mode: bool,
+    pub vision_swap: bool,
+    pub patches_per_batch: usize,
+    pub cpu_patches: usize,
+    pub vision_offload: VisionOffload,
     #[serde(flatten)]
     pub decode: DecodeParameters,
 }
@@ -215,6 +231,10 @@ impl Default for InferenceSettings {
             base_size: 1024,
             image_size: 640,
             crop_mode: true,
+            vision_swap: true,
+            patches_per_batch: 2,
+            vision_offload: VisionOffload::default(),
+            cpu_patches: 0,
             decode: DecodeParameters::default(),
         }
     }
@@ -526,6 +546,10 @@ pub struct InferenceOverride {
     pub base_size: Option<u32>,
     pub image_size: Option<u32>,
     pub crop_mode: Option<bool>,
+    pub vision_swap: Option<bool>,
+    pub patches_per_batch: Option<usize>,
+    pub vision_offload: Option<VisionOffload>,
+    pub cpu_patches: Option<usize>,
     #[serde(flatten)]
     pub decode: DecodeParametersPatch,
 }
@@ -550,6 +574,18 @@ impl std::ops::AddAssign<&InferenceOverride> for InferenceSettings {
         if let Some(crop_mode) = rhs.crop_mode {
             self.crop_mode = crop_mode;
         }
+        if let Some(vision_swap) = rhs.vision_swap {
+            self.vision_swap = vision_swap;
+        }
+        if let Some(patches_per_batch) = rhs.patches_per_batch {
+            self.patches_per_batch = patches_per_batch;
+        }
+        if let Some(vision_offload) = rhs.vision_offload {
+            self.vision_offload = vision_offload;
+        }
+        if let Some(cpu_patches) = rhs.cpu_patches {
+            self.cpu_patches = cpu_patches;
+        }
 
         self.decode += &rhs.decode;
     }
@@ -561,6 +597,10 @@ impl InferenceSettings {
             base_size: self.base_size,
             image_size: self.image_size,
             crop_mode: self.crop_mode,
+            vision_swap: self.vision_swap,
+            patches_per_batch: self.patches_per_batch,
+            cpu_patches: self.cpu_patches,
+            vision_offload: self.vision_offload,
         }
     }
 }
diff --git a/crates/core/src/inference.rs b/crates/core/src/inference.rs
index 2fd10bd..292e644 100644
--- a/crates/core/src/inference.rs
+++ b/crates/core/src/inference.rs
@@ -4,7 +4,7 @@ use image::DynamicImage;
 use serde::{Deserialize, Serialize};
 use tokenizers::Tokenizer;
 
-use crate::{benchmark::Timer, conversation::get_conv_template, sampling::TokenSelectionParams};
+use crate::{benchmark::Timer, conversation::get_conv_template, runtime::VisionOffload, sampling::TokenSelectionParams};
 
 /// Callback used to stream decoded token pieces.
 pub type StreamCallback<'a> = Option<&'a dyn Fn(usize, &[i64])>;
@@ -15,6 +15,31 @@ pub struct VisionSettings {
     pub base_size: u32,
     pub image_size: u32,
     pub crop_mode: bool,
+    /// Enable VRAM swap for vision models (default: true = auto-detect).
+    /// Set to false to force SAM/CLIP on CPU.
+    pub vision_swap: bool,
+    /// Patch batch size for VRAM swap (default: 2).
+    /// Smaller values use less VRAM but are slower.
+    pub patches_per_batch: usize,
+    /// Vision offload strategy (overrides vision_swap when non-Auto).
+    pub vision_offload: VisionOffload,
+    /// Number of patches to process on CPU in sequential mode (default: 0 = all on GPU).
+    /// First `cpu_patches` patches run on CPU alongside global view; the rest use VRAM swap on GPU.
+    pub cpu_patches: usize,
+}
+
+impl Default for VisionSettings {
+    fn default() -> Self {
+        Self {
+            base_size: 1024,
+            image_size: 640,
+            crop_mode: true,
+            vision_swap: true,
+            patches_per_batch: 2,
+            vision_offload: VisionOffload::Auto,
+            cpu_patches: 0,
+        }
+    }
 }
 
 /// Decoding parameters that map directly onto generation options.
diff --git a/crates/core/src/lib.rs b/crates/core/src/lib.rs
index af92747..9943052 100644
--- a/crates/core/src/lib.rs
+++ b/crates/core/src/lib.rs
@@ -11,6 +11,7 @@ pub use inference::{
     DecodeOutcome, DecodeParameters, DecodeParametersPatch, ModelKind, ModelLoadArgs, OcrEngine,
     VisionSettings, normalize_text, render_prompt,
 };
+pub use runtime::{DeviceKind, Precision, VisionOffload, default_dtype_for_device, dtype_from_precision, prepare_device_and_dtype};
 
 // #[cfg(feature = "mkl")]
 // extern crate intel_mkl_src;
diff --git a/crates/core/src/runtime.rs b/crates/core/src/runtime.rs
index b3cb367..593470c 100644
--- a/crates/core/src/runtime.rs
+++ b/crates/core/src/runtime.rs
@@ -19,6 +19,20 @@ pub enum Precision {
     Bf16,
 }
 
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Default, ValueEnum, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum VisionOffload {
+    /// Auto-detect based on VRAM: use Sequential on <6GB, normal on >=6GB.
+    #[default]
+    Auto,
+    /// Hybrid: global view on CPU, patch crops on CUDA (current swap behavior).
+    Sequential,
+    /// Full GPU: load SAM then CLIP sequentially on CUDA for all vision.
+    FullGpu,
+    /// Force all vision on CPU (no VRAM swap).
+    Cpu,
+}
+
 pub fn prepare_device_and_dtype(
     device: DeviceKind,
     precision: Option<Precision>,
diff --git a/crates/core/src/tensor.rs b/crates/core/src/tensor.rs
index d4b9005..8467bc2 100644
--- a/crates/core/src/tensor.rs
+++ b/crates/core/src/tensor.rs
@@ -5,7 +5,16 @@ use candle_core::{DType, Tensor};
 ///
 /// `weight` must be `[vocab, hidden]` and `ids` must be rank-2 `[batch, seq]`. The returned tensor
 /// has shape `[batch, seq, hidden]`.
+///
+/// When `weight` and `ids` reside on different devices, `ids` are temporarily moved to
+/// `weight`'s device for the lookup and the result is moved back to `ids`' original device.
 pub fn gather_token_embeddings(weight: &Tensor, ids: &Tensor) -> Result<Tensor> {
+    let target_device = ids.device();
+    let ids = if !ids.device().same_device(weight.device()) {
+        ids.to_device(weight.device())?
+    } else {
+        ids.clone()
+    };
     ensure!(
         ids.rank() == 2,
         "input ids must have shape [batch, seq], got rank {}",
@@ -14,12 +23,17 @@ pub fn gather_token_embeddings(weight: &Tensor, ids: &Tensor) -> Result<Tensor>
     let (_vocab, hidden) = weight.shape().dims2()?;
     let (batch, seq_len) = ids.shape().dims2()?;
     let ids = if ids.dtype() == DType::I64 {
-        ids.clone()
+        ids
     } else {
         ids.to_dtype(DType::I64)?
     };
     let weight = weight.force_contiguous()?;
     let flat = ids.reshape((batch * seq_len,))?.force_contiguous()?;
     let gathered = weight.index_select(&flat, 0)?;
-    Ok(gathered.reshape((batch, seq_len, hidden))?)
+    let gathered = gathered.reshape((batch, seq_len, hidden))?;
+    if !gathered.device().same_device(target_device) {
+        Ok(gathered.to_device(target_device)?)
+    } else {
+        Ok(gathered)
+    }
 }
diff --git a/crates/dsq-cli/src/main.rs b/crates/dsq-cli/src/main.rs
index afea188..7cbd2d0 100644
--- a/crates/dsq-cli/src/main.rs
+++ b/crates/dsq-cli/src/main.rs
@@ -13,7 +13,8 @@ use deepseek_ocr_dsq_models::{
     AdapterRegistry, AdapterScope, LinearSpec, ModelAdapter, QuantContext,
 };
 use deepseek_ocr_dsq_writer::{
-    encode_bias_values, quantize_q4k, quantize_q6k, quantize_q8_0, DsqWriter, SnapshotMetadata,
+    encode_bias_values, quantize_q2k, quantize_q3k, quantize_q4k, quantize_q6k, quantize_q8_0,
+    DsqWriter, SnapshotMetadata,
 };
 use half::{bf16, f16};
 use indicatif::{ProgressBar, ProgressDrawTarget, ProgressStyle};
@@ -81,6 +82,10 @@ struct StatsArgs {
 enum QuantDTypeArg {
     #[value(name = "Q8_0")]
     Q8_0,
+    #[value(name = "Q2_K")]
+    Q2K,
+    #[value(name = "Q3_K")]
+    Q3K,
     #[value(name = "Q4_K")]
     Q4K,
     #[value(name = "Q6_K")]
@@ -91,6 +96,8 @@ impl QuantDTypeArg {
     fn label(self) -> &'static str {
         match self {
             Self::Q8_0 => "Q8_0",
+            Self::Q2K => "Q2_K",
+            Self::Q3K => "Q3_K",
             Self::Q4K => "Q4_K",
             Self::Q6K => "Q6_K",
         }
@@ -99,6 +106,8 @@ impl QuantDTypeArg {
     fn to_dtype(self) -> DsqTensorDType {
         match self {
             Self::Q8_0 => DsqTensorDType::Q8_0,
+            Self::Q2K => DsqTensorDType::Q2K,
+            Self::Q3K => DsqTensorDType::Q3K,
             Self::Q4K => DsqTensorDType::Q4K,
             Self::Q6K => DsqTensorDType::Q6K,
         }
@@ -617,6 +626,8 @@ fn quantize_spec(
             }
             let qbytes = match selection.dtype {
                 DsqTensorDType::Q8_0 => quantize_q8_0(&weights, spec.out_dim, spec.in_dim)?,
+                DsqTensorDType::Q2K => quantize_q2k(&weights, spec.out_dim, spec.in_dim)?,
+                DsqTensorDType::Q3K => quantize_q3k(&weights, spec.out_dim, spec.in_dim)?,
                 DsqTensorDType::Q4K => quantize_q4k(&weights, spec.out_dim, spec.in_dim)?,
                 DsqTensorDType::Q6K => quantize_q6k(&weights, spec.out_dim, spec.in_dim)?,
                 other => unreachable!("float dtype {other:?} cannot be selected here"),
@@ -991,7 +1002,9 @@ fn select_dtype(primary: DsqTensorDType, in_dim: usize) -> Result<SelectionResul
 
 fn next_fallback_dtype(dtype: DsqTensorDType) -> Option<DsqTensorDType> {
     match dtype {
-        DsqTensorDType::Q6K | DsqTensorDType::Q4K => Some(DsqTensorDType::Q8_0),
+        DsqTensorDType::Q2K | DsqTensorDType::Q3K | DsqTensorDType::Q4K | DsqTensorDType::Q6K => {
+            Some(DsqTensorDType::Q8_0)
+        }
         DsqTensorDType::Q8_0 => None,
         _ => None,
     }
diff --git a/crates/dsq-models/src/adapters/deepseek_ocr.rs b/crates/dsq-models/src/adapters/deepseek_ocr.rs
index 27a1090..fcc5244 100644
--- a/crates/dsq-models/src/adapters/deepseek_ocr.rs
+++ b/crates/dsq-models/src/adapters/deepseek_ocr.rs
@@ -149,6 +149,14 @@ impl ModelAdapter for DeepSeekOcrAdapter {
         }
         match tensor {
             "lm_head.weight" | "model.projector.layers.weight" => Some(DsqTensorDType::Q8_0),
+            // gate_proj and up_proj form the gating pathway — very sensitive to
+            // precision loss.  When the primary dtype is Q2_K, promote them to
+            // Q4_K to preserve the expert routing signal inside each MoE layer.
+            name if ctx.primary == DsqTensorDType::Q2K
+                && (name.contains("gate_proj") || name.contains("up_proj")) =>
+            {
+                Some(DsqTensorDType::Q4K)
+            }
             _ => None,
         }
     }
diff --git a/crates/dsq-runtime/src/lib.rs b/crates/dsq-runtime/src/lib.rs
index 72d3436..8c02c4a 100644
--- a/crates/dsq-runtime/src/lib.rs
+++ b/crates/dsq-runtime/src/lib.rs
@@ -336,7 +336,11 @@ impl QuantizedSnapshot {
         let qweight_bytes = self.reader.tensor_bytes(record)?;
         let bias = self.load_bias(record, name, device)?;
         match record.q_dtype {
-            DsqTensorDType::Q8_0 | DsqTensorDType::Q4K | DsqTensorDType::Q6K => {
+            DsqTensorDType::Q8_0
+            | DsqTensorDType::Q2K
+            | DsqTensorDType::Q3K
+            | DsqTensorDType::Q4K
+            | DsqTensorDType::Q6K => {
                 let ggml_dtype = ggml_from_snapshot_dtype(record.q_dtype)?;
                 let qtensor = qtensor_from_ggml(
                     ggml_dtype,
@@ -580,6 +584,8 @@ impl QuantizedSnapshot {
 fn ggml_from_snapshot_dtype(dtype: DsqTensorDType) -> Result<GgmlDType> {
     match dtype {
         DsqTensorDType::Q8_0 => Ok(GgmlDType::Q8_0),
+        DsqTensorDType::Q2K => Ok(GgmlDType::Q2K),
+        DsqTensorDType::Q3K => Ok(GgmlDType::Q3K),
         DsqTensorDType::Q4K => Ok(GgmlDType::Q4K),
         DsqTensorDType::Q6K => Ok(GgmlDType::Q6K),
         other => bail!("snapshot dtype {:?} does not map to ggml", other),
@@ -588,7 +594,11 @@ fn ggml_from_snapshot_dtype(dtype: DsqTensorDType) -> Result<GgmlDType> {
 
 fn ensure_supported_snapshot_dtype(dtype: DsqTensorDType) -> Result<()> {
     match dtype {
-        DsqTensorDType::Q8_0 | DsqTensorDType::Q4K | DsqTensorDType::Q6K => Ok(()),
+        DsqTensorDType::Q8_0
+        | DsqTensorDType::Q2K
+        | DsqTensorDType::Q3K
+        | DsqTensorDType::Q4K
+        | DsqTensorDType::Q6K => Ok(()),
         DsqTensorDType::F16 | DsqTensorDType::BF16 | DsqTensorDType::F32 => Ok(()),
     }
 }
diff --git a/crates/dsq-writer/src/lib.rs b/crates/dsq-writer/src/lib.rs
index 442dc06..00170aa 100644
--- a/crates/dsq-writer/src/lib.rs
+++ b/crates/dsq-writer/src/lib.rs
@@ -6,7 +6,7 @@ use std::{
     slice,
 };
 
-use candle_core::quantized::k_quants::{BlockQ4K, BlockQ6K, GgmlType as CandleGgmlType};
+use candle_core::quantized::k_quants::{BlockQ2K, BlockQ3K, BlockQ4K, BlockQ6K, GgmlType as CandleGgmlType};
 
 use deepseek_ocr_dsq::{DsqBiasDType, DsqTensorDType};
 use half::{bf16, f16};
@@ -16,6 +16,10 @@ use thiserror::Error;
 const DSQ_MAGIC: &[u8; 7] = b"DSQSNAP";
 const DSQ_VERSION: u32 = 1;
 const Q8_BLOCK: usize = 32;
+const Q2K_BLOCK: usize = 256;
+const Q2K_BLOCK_BYTES: usize = mem::size_of::<BlockQ2K>();
+const Q3K_BLOCK: usize = 256;
+const Q3K_BLOCK_BYTES: usize = mem::size_of::<BlockQ3K>();
 const Q4K_BLOCK: usize = 256;
 const Q4K_BLOCK_BYTES: usize = mem::size_of::<BlockQ4K>(); // 144 bytes per block (K-scale layout)
 const Q6K_BLOCK: usize = 256;
@@ -167,6 +171,116 @@ impl DsqWriter {
         )
     }
 
+    /// Quantize a dense matrix into Q2_K blocks and append it as a tensor record.
+    pub fn add_q2k_tensor(
+        &mut self,
+        name: impl Into<String>,
+        out_dim: usize,
+        in_dim: usize,
+        weights: &[f32],
+        bias: Option<&[f32]>,
+    ) -> Result<()> {
+        let name = name.into();
+        if self.records.iter().any(|rec| rec.name == name) {
+            return Err(DsqWriterError::DuplicateTensor(name.clone()));
+        }
+        if !in_dim.is_multiple_of(Q2K_BLOCK) {
+            return Err(DsqWriterError::InvalidBlock {
+                name: name.clone(),
+                in_dim,
+                block: Q2K_BLOCK,
+            });
+        }
+        let expected = out_dim
+            .checked_mul(in_dim)
+            .ok_or(DsqWriterError::ValueOverflow {
+                what: "tensor elements",
+            })?;
+        if weights.len() != expected {
+            return Err(DsqWriterError::DimensionMismatch {
+                name: name.clone(),
+                expected,
+                found: weights.len(),
+            });
+        }
+        if let Some(bias_vals) = bias {
+            if bias_vals.len() != out_dim {
+                return Err(DsqWriterError::BiasLengthMismatch {
+                    name: name.clone(),
+                    out_dim,
+                    found: bias_vals.len(),
+                });
+            }
+        }
+        let qbytes = quantize_q2k(weights, out_dim, in_dim)?;
+        let bias_bytes = bias.map(encode_bias_values);
+        self.add_quantized_tensor_internal(
+            name,
+            out_dim,
+            in_dim,
+            DsqTensorDType::Q2K,
+            &qbytes,
+            bias_bytes
+                .as_deref()
+                .map(|slice| (slice, DsqBiasDType::F32)),
+        )
+    }
+
+    /// Quantize a dense matrix into Q3_K blocks and append it as a tensor record.
+    pub fn add_q3k_tensor(
+        &mut self,
+        name: impl Into<String>,
+        out_dim: usize,
+        in_dim: usize,
+        weights: &[f32],
+        bias: Option<&[f32]>,
+    ) -> Result<()> {
+        let name = name.into();
+        if self.records.iter().any(|rec| rec.name == name) {
+            return Err(DsqWriterError::DuplicateTensor(name.clone()));
+        }
+        if !in_dim.is_multiple_of(Q3K_BLOCK) {
+            return Err(DsqWriterError::InvalidBlock {
+                name: name.clone(),
+                in_dim,
+                block: Q3K_BLOCK,
+            });
+        }
+        let expected = out_dim
+            .checked_mul(in_dim)
+            .ok_or(DsqWriterError::ValueOverflow {
+                what: "tensor elements",
+            })?;
+        if weights.len() != expected {
+            return Err(DsqWriterError::DimensionMismatch {
+                name: name.clone(),
+                expected,
+                found: weights.len(),
+            });
+        }
+        if let Some(bias_vals) = bias {
+            if bias_vals.len() != out_dim {
+                return Err(DsqWriterError::BiasLengthMismatch {
+                    name: name.clone(),
+                    out_dim,
+                    found: bias_vals.len(),
+                });
+            }
+        }
+        let qbytes = quantize_q3k(weights, out_dim, in_dim)?;
+        let bias_bytes = bias.map(encode_bias_values);
+        self.add_quantized_tensor_internal(
+            name,
+            out_dim,
+            in_dim,
+            DsqTensorDType::Q3K,
+            &qbytes,
+            bias_bytes
+                .as_deref()
+                .map(|slice| (slice, DsqBiasDType::F32)),
+        )
+    }
+
     /// Quantize a dense matrix into Q4_K blocks and append it as a tensor record.
     pub fn add_q4k_tensor(
         &mut self,
@@ -597,6 +711,72 @@ pub fn quantize_q8_0(weights: &[f32], rows: usize, cols: usize) -> Result<Vec<u8
     Ok(result)
 }
 
+pub fn quantize_q2k(weights: &[f32], rows: usize, cols: usize) -> Result<Vec<u8>> {
+    if !cols.is_multiple_of(Q2K_BLOCK) {
+        return Err(DsqWriterError::InvalidBlock {
+            name: "quantize_q2k".into(),
+            in_dim: cols,
+            block: Q2K_BLOCK,
+        });
+    }
+    if weights.len() != rows * cols {
+        return Err(DsqWriterError::DimensionMismatch {
+            name: "quantize_q2k".into(),
+            expected: rows * cols,
+            found: weights.len(),
+        });
+    }
+    let blocks_per_row = cols / Q2K_BLOCK;
+    let total_blocks = rows
+        .checked_mul(blocks_per_row)
+        .ok_or(DsqWriterError::ValueOverflow { what: "q2k blocks" })?;
+    let mut result = Vec::with_capacity(total_blocks * Q2K_BLOCK_BYTES);
+    for row in 0..rows {
+        let start = row * cols;
+        let row_slice = &weights[start..start + cols];
+        let mut blocks = vec![<BlockQ2K as CandleGgmlType>::zeros(); blocks_per_row];
+        <BlockQ2K as CandleGgmlType>::from_float(row_slice, &mut blocks);
+        let bytes = unsafe {
+            slice::from_raw_parts(blocks.as_ptr() as *const u8, blocks.len() * Q2K_BLOCK_BYTES)
+        };
+        result.extend_from_slice(bytes);
+    }
+    Ok(result)
+}
+
+pub fn quantize_q3k(weights: &[f32], rows: usize, cols: usize) -> Result<Vec<u8>> {
+    if !cols.is_multiple_of(Q3K_BLOCK) {
+        return Err(DsqWriterError::InvalidBlock {
+            name: "quantize_q3k".into(),
+            in_dim: cols,
+            block: Q3K_BLOCK,
+        });
+    }
+    if weights.len() != rows * cols {
+        return Err(DsqWriterError::DimensionMismatch {
+            name: "quantize_q3k".into(),
+            expected: rows * cols,
+            found: weights.len(),
+        });
+    }
+    let blocks_per_row = cols / Q3K_BLOCK;
+    let total_blocks = rows
+        .checked_mul(blocks_per_row)
+        .ok_or(DsqWriterError::ValueOverflow { what: "q3k blocks" })?;
+    let mut result = Vec::with_capacity(total_blocks * Q3K_BLOCK_BYTES);
+    for row in 0..rows {
+        let start = row * cols;
+        let row_slice = &weights[start..start + cols];
+        let mut blocks = vec![<BlockQ3K as CandleGgmlType>::zeros(); blocks_per_row];
+        <BlockQ3K as CandleGgmlType>::from_float(row_slice, &mut blocks);
+        let bytes = unsafe {
+            slice::from_raw_parts(blocks.as_ptr() as *const u8, blocks.len() * Q3K_BLOCK_BYTES)
+        };
+        result.extend_from_slice(bytes);
+    }
+    Ok(result)
+}
+
 pub fn quantize_q4k(weights: &[f32], rows: usize, cols: usize) -> Result<Vec<u8>> {
     if !cols.is_multiple_of(Q4K_BLOCK) {
         return Err(DsqWriterError::InvalidBlock {
@@ -688,6 +868,8 @@ fn expected_qbyte_len(
     let blocks_per_row = in_dim / block;
     let per_block = match dtype {
         DsqTensorDType::Q8_0 => Q8_BLOCK_BYTES,
+        DsqTensorDType::Q2K => Q2K_BLOCK_BYTES,
+        DsqTensorDType::Q3K => Q3K_BLOCK_BYTES,
         DsqTensorDType::Q4K => Q4K_BLOCK_BYTES,
         DsqTensorDType::Q6K => Q6K_BLOCK_BYTES,
         other => unreachable!("expected quantized dtype, got {other:?}"),
diff --git a/crates/dsq/src/lib.rs b/crates/dsq/src/lib.rs
index dc4f3fa..279c85a 100644
--- a/crates/dsq/src/lib.rs
+++ b/crates/dsq/src/lib.rs
@@ -50,6 +50,8 @@ pub struct DsqHeader {
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
 pub enum DsqTensorDType {
     Q8_0,
+    Q2K,
+    Q3K,
     Q4K,
     Q6K,
     F16,
@@ -61,6 +63,8 @@ impl DsqTensorDType {
     pub fn as_u32(self) -> u32 {
         match self {
             Self::Q8_0 => 8,
+            Self::Q2K => 10,
+            Self::Q3K => 11,
             Self::Q4K => 12,
             Self::Q6K => 14,
             Self::F16 => 1,
@@ -72,6 +76,8 @@ impl DsqTensorDType {
     pub fn block_size(self) -> Option<usize> {
         match self {
             Self::Q8_0 => Some(32),
+            Self::Q2K => Some(256),
+            Self::Q3K => Some(256),
             Self::Q4K => Some(256),
             Self::Q6K => Some(256),
             Self::F16 | Self::BF16 | Self::F32 => None,
@@ -97,6 +103,8 @@ impl TryFrom<u32> for DsqTensorDType {
     fn try_from(value: u32) -> Result<Self> {
         match value {
             8 => Ok(Self::Q8_0),
+            10 => Ok(Self::Q2K),
+            11 => Ok(Self::Q3K),
             12 => Ok(Self::Q4K),
             14 => Ok(Self::Q6K),
             1 => Ok(Self::F16),
@@ -113,6 +121,8 @@ impl fmt::Display for DsqTensorDType {
     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
         match self {
             Self::Q8_0 => f.write_str("Q8_0"),
+            Self::Q2K => f.write_str("Q2_K"),
+            Self::Q3K => f.write_str("Q3_K"),
             Self::Q4K => f.write_str("Q4_K"),
             Self::Q6K => f.write_str("Q6_K"),
             Self::F16 => f.write_str("F16"),
diff --git a/crates/infer-deepseek/src/model/mod.rs b/crates/infer-deepseek/src/model/mod.rs
index 961ff25..950d2e6 100644
--- a/crates/infer-deepseek/src/model/mod.rs
+++ b/crates/infer-deepseek/src/model/mod.rs
@@ -1,4 +1,5 @@
 use std::{
+    cell::UnsafeCell,
     convert::TryFrom,
     path::{Path, PathBuf},
     sync::Arc,
@@ -13,6 +14,9 @@ use rayon::prelude::*;
 use tokenizers::Tokenizer;
 use tracing::{info, trace};
 
+mod swap;
+use swap::{SequentialVramSwap, should_use_vram_swap};
+
 use crate::{
     config::{DeepseekOcrConfig, ProjectorConfig, load_ocr_config},
     quant_snapshot::{LinearSpec, QuantizedSnapshot, SnapshotLinear, SnapshotLoadPlan},
@@ -35,6 +39,7 @@ use deepseek_ocr_core::{
         DecodeOutcome, DecodeParameters, ModelKind, ModelLoadArgs, OcrEngine, VisionSettings,
         normalize_text,
     },
+    runtime::VisionOffload,
     sampling::{TokenSelectionParams, init_rng, select_token_id},
 };
 
@@ -485,7 +490,7 @@ impl ImageProjector {
 /// language model. For now it wires the language stack so we can exercise text-only inference.
 pub struct DeepseekOcrModel {
     cfg: Arc<DeepseekOcrConfig>,
-    language: DeepseekLanguageModel,
+    language: UnsafeCell<DeepseekLanguageModel>,
     projector_cfg: Arc<ProjectorConfig>,
     projector: ImageProjector,
     projector_f32: Option<ImageProjector>,
@@ -525,8 +530,10 @@ struct VisionModules {
 
 struct VisionContext<'a> {
     projector: &'a ImageProjector,
-    vision: &'a VisionModules,
+    sam: &'a SamBackbone,
+    clip: &'a ClipVisionModel,
     device: &'a Device,
+    vision_device: Device,
     dtype: DType,
     parallel: bool,
 }
@@ -557,20 +564,49 @@ fn prepare_image_tensor_for_device(
 impl<'a> VisionContext<'a> {
     fn new_with_dtype(
         model: &'a DeepseekOcrModel,
-        vision: &'a VisionModules,
+        sam: &'a SamBackbone,
+        clip: &'a ClipVisionModel,
         projector: &'a ImageProjector,
         dtype: DType,
     ) -> Self {
-        let parallel = matches!(model.device(), Device::Cpu);
+        let device = model.device();
+        let parallel = matches!(device, Device::Cpu);
+        let vision_device = if device.is_cuda() {
+            Device::Cpu
+        } else {
+            device.clone()
+        };
         Self {
             projector,
-            vision,
-            device: model.device(),
+            sam,
+            clip,
+            device,
+            vision_device,
             dtype,
             parallel,
         }
     }
 
+    /// Create a VisionContext with an explicit vision device (e.g. CUDA for swapped models).
+    fn new_with_explicit_device(
+        model: &'a DeepseekOcrModel,
+        sam: &'a SamBackbone,
+        clip: &'a ClipVisionModel,
+        projector: &'a ImageProjector,
+        dtype: DType,
+        vision_device: Device,
+    ) -> Self {
+        Self {
+            projector,
+            sam,
+            clip,
+            device: model.device(),
+            vision_device,
+            dtype,
+            parallel: false,
+        }
+    }
+
     fn hidden_size(&self) -> usize {
         self.projector.hidden_size()
     }
@@ -584,7 +620,7 @@ impl<'a> VisionContext<'a> {
     }
 
     fn prepare_image_tensor(&self, tensor: &Tensor) -> Result<Tensor> {
-        prepare_image_tensor_for_device(tensor, self.device, self.dtype)
+        prepare_image_tensor_for_device(tensor, &self.vision_device, self.dtype)
     }
 
     fn append_row_breaks(&self, grid: Tensor, newline: &Tensor) -> Result<Tensor> {
@@ -758,12 +794,10 @@ impl<'a> VisionContext<'a> {
             .prepare_image_tensor(input.global)
             .context("invalid global image tensor")?;
         let sam_global = self
-            .vision
             .sam
             .forward(&global)
             .context("sam forward (global)")?;
         let clip_global = self
-            .vision
             .clip
             .forward(&global, Some(&sam_global))
             .context("clip forward (global)")?;
@@ -772,6 +806,11 @@ impl<'a> VisionContext<'a> {
             .context("concat global clip+sam tokens")?
             .contiguous()
             .context("global pre tokens not contiguous")?;
+        let global_pre = if !global_pre.device().same_device(self.device) {
+            global_pre.to_device(self.device)?
+        } else {
+            global_pre
+        };
         let global_post = self
             .projector
             .project(&global_pre)
@@ -851,22 +890,29 @@ impl<'a> VisionContext<'a> {
         self.process_patch_batch(&chunk)
     }
 
-    fn process_patch_batch(&self, batch: &Tensor) -> Result<(Tensor, Tensor)> {
+    /// Process a single image through SAM → CLIP → projector.
+    /// Used by `process_patch_batch` to avoid batching multiple images through
+    /// SAM at once, which would multiply activation memory and cause OOM on
+    /// low-VRAM GPUs.
+    fn process_single_image(&self, image: &Tensor) -> Result<(Tensor, Tensor)> {
         let sam_local = self
-            .vision
             .sam
-            .forward(batch)
+            .forward(image)
             .context("sam forward (local)")?;
         let clip_local = self
-            .vision
             .clip
-            .forward(batch, Some(&sam_local))
+            .forward(image, Some(&sam_local))
             .context("clip forward (local)")?;
         let local_pre = self
             .build_clip_sam_tokens(&clip_local, &sam_local)
             .context("concat local clip+sam tokens")?
             .contiguous()
             .context("local pre tokens not contiguous")?;
+        let local_pre = if !local_pre.device().same_device(self.device) {
+            local_pre.to_device(self.device)?
+        } else {
+            local_pre
+        };
         let local_post = self
             .projector
             .project(&local_pre)
@@ -876,6 +922,31 @@ impl<'a> VisionContext<'a> {
         Ok((local_pre, local_post))
     }
 
+    fn process_patch_batch(&self, batch: &Tensor) -> Result<(Tensor, Tensor)> {
+        let batch_size = batch.shape().dims4().map(|d| d.0).unwrap_or(0);
+        if batch_size <= 1 {
+            return self.process_single_image(batch);
+        }
+        // Process each patch individually to limit peak SAM activation memory
+        let chunks = batch.chunk(batch_size, 0)?;
+        let results: Result<Vec<_>> = chunks
+            .into_iter()
+            .map(|chunk| self.process_single_image(&chunk))
+            .collect();
+        let (pre_list, post_list): (Vec<_>, Vec<_>) = results?
+            .into_iter()
+            .unzip();
+        let pre_refs: Vec<_> = pre_list.iter().collect();
+        let post_refs: Vec<_> = post_list.iter().collect();
+        let local_pre = Tensor::cat(&pre_refs, 0)?
+            .contiguous()
+            .context("batched local pre tokens not contiguous")?;
+        let local_post = Tensor::cat(&post_refs, 0)?
+            .contiguous()
+            .context("batched local post tokens not contiguous")?;
+        Ok((local_pre, local_post))
+    }
+
     fn assemble_artifacts(
         &self,
         global_pre: Tensor,
@@ -990,7 +1061,7 @@ impl DeepseekOcrModel {
         .context("failed to load language model")?;
 
         let low_precision = matches!(dtype, DType::F16 | DType::BF16);
-        if low_precision {
+        if low_precision && !device.is_cuda() {
             let vb_f32_lang = unsafe {
                 VarBuilder::from_mmaped_safetensors(
                     &[resolved_weights.as_path()],
@@ -1036,7 +1107,7 @@ impl DeepseekOcrModel {
         let projector = ImageProjector::load(&vb, projector_cfg.as_ref(), snapshot.as_deref())
             .context("failed to load image projector")?;
         let low_precision = matches!(dtype, DType::F16 | DType::BF16);
-        let (projector_f32, vision_f32, vision_ocr2_f32) = if low_precision {
+        let (projector_f32, vision_f32, vision_ocr2_f32) = if low_precision && !device.is_cuda() {
             let vb_f32 = unsafe {
                 VarBuilder::from_mmaped_safetensors(
                     &[resolved_weights.as_path()],
@@ -1074,10 +1145,34 @@ impl DeepseekOcrModel {
         };
         let vision = match variant {
             OcrVariant::Ocr1 => {
-                let sam = SamBackbone::new(cfg.as_ref(), &vb.pp("model").pp("sam_model"))
-                    .context("failed to load SAM backbone")?;
-                let clip = ClipVisionModel::load(cfg.as_ref(), &vb.pp("model").pp("vision_model"))
-                    .context("failed to load CLIP vision model")?;
+                let (sam, clip) = if device.is_cuda() {
+                    let vb_cpu = unsafe {
+                        VarBuilder::from_mmaped_safetensors(
+                            &[resolved_weights.as_path()],
+                            dtype,
+                            &Device::Cpu,
+                        )
+                    }
+                    .with_context(|| {
+                        format!(
+                            "failed to mmap cpu weights for vision at {}",
+                            resolved_weights.display()
+                        )
+                    })?;
+                    let sam = SamBackbone::new(cfg.as_ref(), &vb_cpu.pp("model").pp("sam_model"))
+                        .context("failed to load SAM backbone on CPU")?;
+                    let clip =
+                        ClipVisionModel::load(cfg.as_ref(), &vb_cpu.pp("model").pp("vision_model"))
+                            .context("failed to load CLIP vision model on CPU")?;
+                    (sam, clip)
+                } else {
+                    let sam = SamBackbone::new(cfg.as_ref(), &vb.pp("model").pp("sam_model"))
+                        .context("failed to load SAM backbone")?;
+                    let clip =
+                        ClipVisionModel::load(cfg.as_ref(), &vb.pp("model").pp("vision_model"))
+                            .context("failed to load CLIP vision model")?;
+                    (sam, clip)
+                };
                 VisionBackend::Ocr1(Box::new(VisionModules { sam, clip }))
             }
             OcrVariant::Ocr2 => VisionBackend::Ocr2(Box::new(
@@ -1088,9 +1183,16 @@ impl DeepseekOcrModel {
         // Log quantization summary after all quantizable modules (language + projector) are loaded.
         QuantizationState::global().log_summary(&device);
 
+        // On CUDA, keep the token embedding on CPU to free VRAM for inference activations.
+        if device.is_cuda() {
+            language
+                .move_token_embedding_to(&Device::Cpu)
+                .context("failed to move token embedding to CPU")?;
+        }
+
         Ok(Self {
             cfg,
-            language,
+            language: UnsafeCell::new(language),
             projector_cfg,
             projector,
             projector_f32,
@@ -1126,12 +1228,23 @@ impl DeepseekOcrModel {
 
     /// Borrow the language-only component.
     pub fn language_model(&self) -> &DeepseekLanguageModel {
-        &self.language
+        // SAFETY: language is only mutated during controlled swap operations
+        // (swap_language_for_vision/restore_language) which are single-threaded
+        // and never overlap with read accesses.
+        unsafe { &*self.language.get() }
+    }
+
+    /// Mutably borrow the language model (for device swap during FullGpu vision).
+    /// # Safety
+    /// Caller must ensure no concurrent access via language_model().
+    pub unsafe fn language_model_mut(&self) -> &mut DeepseekLanguageModel {
+        // SAFETY: caller must ensure no concurrent mutable access
+        unsafe { &mut *self.language.get() }
     }
 
     /// Whether flash attention is enabled for the underlying decoder.
     pub fn flash_attention_enabled(&self) -> bool {
-        self.language.flash_attention_enabled()
+        self.language_model().flash_attention_enabled()
     }
 
     /// Access the projector configuration.
@@ -1141,21 +1254,22 @@ impl DeepseekOcrModel {
 
     /// Construct a fresh dynamic cache sized for this model.
     pub fn new_cache(&self) -> DynamicCache {
-        let layers = self.language.transformer_weights().layers.len();
+        let layers = self.language_model().transformer_weights().layers.len();
         DynamicCache::with_num_layers(layers)
     }
 
     /// Construct a fresh dynamic cache sized for this model, matching the model dtype.
     pub fn new_cache_for_dtype(&self, dtype: DType) -> Result<DynamicCache> {
-        let layers = self.language.transformer_weights().layers.len();
+        let lm = self.language_model();
+        let layers = lm.transformer_weights().layers.len();
         let mut cache = DynamicCache::with_num_layers(layers);
         // Pre-seed a zero-length cache entry per layer so cache dtype is deterministic.
         // Low-precision models keep cache storage in f32 to reduce accumulation drift.
         let store_dtype = cache_store_dtype(self.dtype, dtype);
         for layer in 0..layers {
-            let heads = self.language.config().num_attention_heads;
-            let head_dim = self.language.config().hidden_size / heads;
-            let v_head_dim = self.language.config().v_head_dim.unwrap_or(head_dim);
+            let heads = lm.config().num_attention_heads;
+            let head_dim = lm.config().hidden_size / heads;
+            let v_head_dim = lm.config().v_head_dim.unwrap_or(head_dim);
             let device = self.device.clone();
             let key_t =
                 Tensor::zeros((1, heads, head_dim, 0), store_dtype, &device)?.contiguous()?;
@@ -1168,7 +1282,23 @@ impl DeepseekOcrModel {
 
     /// Helper to guard prompt-scoped cache state.
     pub fn prompt_guard<'a>(&'a self, cache: &'a mut DynamicCache) -> PromptCacheGuard<'a> {
-        self.language.prompt_guard(cache)
+        self.language_model().prompt_guard(cache)
+    }
+
+    /// Move the language model to `target` device in-place, freeing its old device memory.
+    ///
+    /// # Safety
+    ///
+    /// No code may access `language_model()` during or between calls to this method.
+    /// Typically called immediately before and after vision processing on a different device.
+    pub unsafe fn move_language_to_device(&self, target: &Device) -> Result<()> {
+        // SAFETY: caller must ensure no concurrent access to self.language
+        let old = unsafe { std::ptr::read(self.language.get()) };
+        let moved = old.to_device(target)?;
+        // `old` is implicitly freed here as `read` gives us ownership.
+        // Overwrite the stale bytes with the new value.
+        unsafe { std::ptr::write(self.language.get(), moved) };
+        Ok(())
     }
 
     #[doc(hidden)]
@@ -1209,7 +1339,7 @@ impl DeepseekOcrModel {
             Some(t) => t.clone(),
             None => {
                 let ids = input_ids.expect("input_ids validity checked above");
-                self.language.embed_tokens(ids)?
+                self.language_model().embed_tokens(ids)?
             }
         };
 
@@ -1238,7 +1368,7 @@ impl DeepseekOcrModel {
             "failed to cast language input embeddings",
         )?;
 
-        let lm_out = self.language.forward(
+        let lm_out = self.language_model().forward(
             None,
             Some(&embeddings),
             attention_mask,
@@ -1273,44 +1403,41 @@ impl DeepseekOcrModel {
         )
     }
 
+    /// Compute image embeddings using the model's stored vision modules.
+    /// Uses auto-detection for VRAM swap (same as default VisionSettings).
     pub fn compute_image_embeddings(
         &self,
         inputs: &[Option<VisionInput<'_>>],
     ) -> Result<Vec<Tensor>> {
+        self.compute_image_embeddings_impl(inputs, &VisionSettings::default())
+    }
+
+    /// Internal implementation with explicit VisionSettings.
+    fn compute_image_embeddings_impl(
+        &self,
+        inputs: &[Option<VisionInput<'_>>],
+        vision: &VisionSettings,
+    ) -> Result<Vec<Tensor>> {
+        let vision_swap = match vision.vision_offload {
+            VisionOffload::Auto => vision.vision_swap,
+            VisionOffload::Sequential | VisionOffload::FullGpu => true,
+            VisionOffload::Cpu => false,
+        };
         match &self.vision {
             VisionBackend::Ocr1(_vision) => {
+                if vision.vision_offload == VisionOffload::FullGpu {
+                    info!("Full GPU mode for vision — loading SAM then CLIP sequentially on CUDA");
+                    return self.compute_image_embeddings_full_gpu(inputs, vision.patches_per_batch);
+                }
+                if vision_swap && should_use_vram_swap(self.device()) {
+                    info!("Low VRAM device — sequential swap mode for vision");
+                    return self.compute_image_embeddings_with_swap(inputs, vision.cpu_patches, vision.patches_per_batch);
+                }
+                let vision = self.vision_modules().context("vision modules missing")?;
                 let compute_dtype = low_precision_compute_dtype(self.dtype);
-                let vision_native = self.vision_modules().context("vision modules missing")?;
-                let vision = select_f32(compute_dtype, vision_native, self.vision_modules_f32());
+                let vision = select_f32(compute_dtype, vision, self.vision_modules_f32());
                 let projector = self.projector_for_dtype(compute_dtype);
-                let ctx = VisionContext::new_with_dtype(self, vision, projector, compute_dtype);
-                let hidden = ctx.hidden_size();
-                let device = ctx.device();
-                if ctx.parallel_enabled() {
-                    inputs
-                        .par_iter()
-                        .map(|input| {
-                            if let Some(vision_input) = input {
-                                ctx.process_input(vision_input)
-                            } else {
-                                Tensor::zeros((0, hidden), compute_dtype, device)
-                                    .map_err(Into::into)
-                            }
-                        })
-                        .collect::<Result<Vec<_>>>()
-                } else {
-                    inputs
-                        .iter()
-                        .map(|input| {
-                            if let Some(vision_input) = input {
-                                ctx.process_input(vision_input)
-                            } else {
-                                Tensor::zeros((0, hidden), compute_dtype, device)
-                                    .map_err(Into::into)
-                            }
-                        })
-                        .collect::<Result<Vec<_>>>()
-                }
+                self.compute_image_embeddings_with(inputs, &vision.sam, &vision.clip, projector, &Device::Cpu)
             }
             VisionBackend::Ocr2(vision) => {
                 let hidden = self.projector.hidden_size();
@@ -1376,6 +1503,455 @@ impl DeepseekOcrModel {
         }
     }
 
+    /// Compute image embeddings using an explicit vision module and vision device.
+    fn compute_image_embeddings_with(
+        &self,
+        inputs: &[Option<VisionInput<'_>>],
+        sam: &SamBackbone,
+        clip: &ClipVisionModel,
+        projector: &ImageProjector,
+        vision_device: &Device,
+    ) -> Result<Vec<Tensor>> {
+        let compute_dtype = low_precision_compute_dtype(self.dtype);
+        let ctx = VisionContext::new_with_explicit_device(self, sam, clip, projector, compute_dtype, vision_device.clone());
+        let hidden = ctx.hidden_size();
+        let device = ctx.device();
+        if ctx.parallel_enabled() {
+            inputs
+                .par_iter()
+                .map(|input| {
+                    if let Some(vision_input) = input {
+                        ctx.process_input(vision_input)
+                    } else {
+                        Tensor::zeros((0, hidden), compute_dtype, device)
+                            .map_err(Into::into)
+                    }
+                })
+                .collect::<Result<Vec<_>>>()
+        } else {
+            inputs
+                .iter()
+                .map(|input| {
+                    if let Some(vision_input) = input {
+                        ctx.process_input(vision_input)
+                    } else {
+                        Tensor::zeros((0, hidden), compute_dtype, device)
+                            .map_err(Into::into)
+                    }
+                })
+                .collect::<Result<Vec<_>>>()
+        }
+    }
+
+    /// Compute image embeddings using hybrid swap: global on CPU, patches on CUDA.
+    ///
+    /// The global view runs on CPU (identical to CPU path); patch crops run on
+    /// CUDA via SequentialVramSwap. Both execute in parallel via scoped threads
+    /// since they use separate hardware (CPU vs GPU).
+    fn compute_image_embeddings_with_swap(
+        &self,
+        inputs: &[Option<VisionInput<'_>>],
+        cpu_patches: usize,
+        chunk_size: usize,
+    ) -> Result<Vec<Tensor>> {
+        let vision = self.vision_modules().context("vision modules missing")?;
+        let compute_dtype = low_precision_compute_dtype(self.dtype);
+        let projector = self.projector_for_dtype(compute_dtype);
+        let device = self.device();
+        let hidden_size = projector.hidden_size();
+        let newline = projector.image_newline_token(compute_dtype, device)?;
+        let view_sep = projector.view_separator_token(compute_dtype, device)?;
+
+        let mut results = Vec::with_capacity(inputs.len());
+
+        for input in inputs {
+            let result = match input {
+                Some(vi) => {
+                    let global = prepare_image_tensor_for_device(
+                        vi.global, &Device::Cpu, compute_dtype,
+                    )?;
+                    let patches_data = vi.patches.map(|p| {
+                        prepare_image_tensor_for_device(p, &Device::Cpu, compute_dtype)
+                    }).transpose()?;
+
+                    let wp = self.weights_path.clone();
+                    let cfg = self.cfg.clone();
+                    let dt = self.dtype;
+
+                    // Split patches: first cpu_patches on CPU, rest on GPU
+                    let patches_split = patches_data.as_ref().map(|p| -> Result<_> {
+                        let (batch, _, _, _) = p.shape().dims4()?;
+                        let cpu_n = cpu_patches.min(batch);
+                        Ok((cpu_n, batch - cpu_n))
+                    }).transpose()?;
+
+                    // Parallelize CPU tasks (global + cpu patches) and GPU patches via scoped threads.
+                    let (combined_global_cpu, cpu_pre_tokens, gpu_projected) = std::thread::scope(|s| {
+                        // Thread 1 (CPU): global SAM+CLIP + cpu_patches SAM+CLIP
+                        let cpu_handle = s.spawn(|| -> Result<(Tensor, Vec<Tensor>)> {
+                            let sg = vision
+                                .sam
+                                .forward(&global)
+                                .context("global SAM forward (CPU)")?;
+                            let cg = vision
+                                .clip
+                                .forward(&global, Some(&sg))
+                                .context("global CLIP forward (CPU)")?;
+                            let (_, clip_seq, _) = cg.shape().dims3()?;
+                            let ct = cg
+                                .narrow(D::Minus2, 1, clip_seq - 1)?
+                                .contiguous()?;
+                            let (_, sc, sh, sw) = sg.shape().dims4()?;
+                            let st = sg
+                                .reshape((1, sc, sh * sw))?
+                                .transpose(1, 2)?
+                                .contiguous()?;
+                            let combined_global = Tensor::cat(&[ct, st], D::Minus1)
+                                .context("cat global combined")?;
+
+                            // Process cpu_patches on CPU
+                            let mut cpu_pres = Vec::new();
+                            if let Some((cpu_n, _)) = patches_split {
+                                if cpu_n > 0 {
+                                    let p = patches_data.as_ref().unwrap();
+                                    let cpu_p = p.narrow(0, 0, cpu_n)?;
+                                    for i in 0..cpu_n {
+                                        let patch = cpu_p.narrow(0, i, 1)?.contiguous()?;
+                                        let ss = vision.sam.forward(&patch)
+                                            .context("cpu patch SAM forward")?;
+                                        let cc = vision.clip.forward(&patch, Some(&ss))
+                                            .context("cpu patch CLIP forward")?;
+                                        let (_, cs, _) = cc.shape().dims3()?;
+                                        let ct = cc.narrow(D::Minus2, 1, cs - 1)?.contiguous()?;
+                                        let (_, sn, sh, sw) = ss.shape().dims4()?;
+                                        let st = ss
+                                            .reshape((1, sn, sh * sw))?
+                                            .transpose(1, 2)?
+                                            .contiguous()?;
+                                        cpu_pres.push(Tensor::cat(&[ct, st], D::Minus1)?);
+                                    }
+                                }
+                            }
+                            Ok((combined_global, cpu_pres))
+                        });
+
+                        // Thread 2 (GPU): gpu_patches via SequentialVramSwap → projected
+                        let gpu_handle = s.spawn(|| -> Result<Option<Vec<Tensor>>> {
+                            let Some((cpu_n, gpu_n)) = patches_split else {
+                                return Ok(None);
+                            };
+                            if gpu_n == 0 {
+                                return Ok(None);
+                            }
+                            let patches = patches_data.as_ref().unwrap();
+                            let gpu_p = patches.narrow(0, cpu_n, gpu_n)?;
+                            let gpu_batch = gpu_n;
+                            let swap = SequentialVramSwap::new(&wp, cfg, dt, device);
+                            let num = (gpu_batch + chunk_size - 1) / chunk_size;
+                            let pcs = gpu_p.chunk(num, 0)?;
+
+                            let sam_cuda = swap
+                                .load_sam_on_cuda()
+                                .context("load SAM on CUDA for patches")?;
+                            let mut sam_cpu = Vec::with_capacity(num);
+                            for c in &pcs {
+                                let cc = c.to_device(device)?;
+                                let o = sam_cuda
+                                    .forward(&cc)
+                                    .context("patch SAM forward (CUDA)")?;
+                                sam_cpu.push(o.to_device(&Device::Cpu)?.contiguous()?);
+                            }
+                            drop(sam_cuda);
+
+                            let clip_cuda = swap
+                                .load_clip_on_cuda()
+                                .context("load CLIP on CUDA for patches")?;
+                            let mut all_proj = Vec::with_capacity(num);
+                            for (pc, sc) in pcs.iter().zip(sam_cpu.iter()) {
+                                let cc = pc.to_device(device)?;
+                                let st = sc.to_device(device)?;
+                                let co = clip_cuda
+                                    .forward(&cc, Some(&st))
+                                    .context("patch CLIP forward (CUDA)")?;
+                                let (_, cseq, _) = co.shape().dims3()?;
+                                let ct = co
+                                    .narrow(D::Minus2, 1, cseq - 1)?
+                                    .contiguous()?;
+                                let (_, scn, sh, sw) = st.shape().dims4()?;
+                                let sts = st
+                                    .reshape((pc.dims()[0], scn, sh * sw))?
+                                    .transpose(1, 2)?
+                                    .contiguous()?;
+                                let cb = Tensor::cat(&[ct, sts], D::Minus1)?;
+                                all_proj.push(
+                                    projector.project(&cb).context("proj forward (patch)")?,
+                                );
+                            }
+                            drop(clip_cuda);
+                            Ok(Some(all_proj))
+                        });
+
+                        let (combined, cpu_pres) = cpu_handle.join().unwrap()?;
+                        let gpu_res = gpu_handle.join().unwrap()?;
+                        Ok::<_, anyhow::Error>((combined, cpu_pres, gpu_res))
+                    })?;
+                    device.synchronize()?;
+
+                    // Project cpu_pres on CUDA and merge with gpu projected tokens
+                    let mut all_projected: Vec<Tensor> = Vec::new();
+                    for cp in &cpu_pre_tokens {
+                        let cp_cuda = cp.to_device(device)?;
+                        all_projected.push(
+                            projector.project(&cp_cuda).context("proj forward (cpu patch)")?,
+                        );
+                    }
+                    if let Some(gpu_proj) = gpu_projected {
+                        all_projected.extend(gpu_proj);
+                    }
+
+                    // Format local grid if any patches were processed
+                    let local_tokens_opt = if !all_projected.is_empty() {
+                        let refs: Vec<&Tensor> = all_projected.iter().collect();
+                        let proj = Tensor::cat(&refs, 0)?;
+                        let total_batch = if let Some((cpu_n, gpu_n)) = patches_split {
+                            cpu_n + gpu_n
+                        } else {
+                            0
+                        };
+                        let cs = vi.crop_shape.unwrap_or((1, total_batch));
+                        let (wc, hc) = cs;
+                        let nl = cast_dtype(&newline, proj.dtype(), "local nl")?;
+                        let (np, seq, hd) = proj.shape().dims3()?;
+                        ensure!(np == wc * hc);
+                        let side = (seq as f64).sqrt() as usize;
+                        ensure!(side * side == seq);
+                        let grid = proj
+                            .reshape((hc, wc, side, side, hd))?
+                            .permute((0, 2, 1, 3, 4))?
+                            .reshape((hc * side, wc * side, hd))?
+                            .contiguous()?;
+                        let (rows, cols, _) = grid.shape().dims3()?;
+                        let nle = nl
+                            .reshape((1, 1, hd))?
+                            .expand((rows, 1, hd))?
+                            .contiguous()?;
+                        let wb = Tensor::cat(&[grid, nle], 1)?;
+                        Some(wb.reshape((rows * (cols + 1), hd))?)
+                    } else {
+                        None
+                    };
+
+                    // Project + format global
+                    let combined_cuda = combined_global_cpu.to_device(device)?;
+                    let projected_global = projector
+                        .project(&combined_cuda)
+                        .context("projector forward (global)")?;
+                    let global_tokens = {
+                        let nl = cast_dtype(&newline, projected_global.dtype(), "nl")?;
+                        let (bs, seq, hd) = projected_global.shape().dims3()?;
+                        ensure!(bs == 1, "global batch must be 1, got {bs}");
+                        let side = (seq as f64).sqrt() as usize;
+                        ensure!(side * side == seq, "global tokens {seq} not square");
+                        let grid = projected_global
+                            .get(0)?
+                            .reshape((side, side, hd))?
+                            .contiguous()?;
+                        let nle = nl
+                            .reshape((1, 1, hd))?
+                            .expand((side, 1, hd))?
+                            .contiguous()?;
+                        Tensor::cat(&[grid, nle], 1)?
+                            .reshape((side * (side + 1), hd))?
+                    };
+
+                    // Assemble final token sequence
+                    let target_dtype = global_tokens.dtype();
+                    let mut segments = Vec::new();
+                    if let Some(lt) = local_tokens_opt {
+                        segments.push(cast_dtype_owned(
+                            lt, target_dtype, "local dtype cast",
+                        )?);
+                    }
+                    segments.push(global_tokens);
+                    let vst = view_sep.reshape((1, hidden_size))?.contiguous()?;
+                    segments.push(cast_dtype_owned(vst, target_dtype, "sep cast")?);
+                    Tensor::cat(&segments, 0)?
+                }
+                None => Tensor::zeros((0, hidden_size), compute_dtype, device)?,
+            };
+            results.push(result);
+        }
+
+        info!("Image embeddings computed via parallel global CPU + patch CUDA");
+        Ok(results)
+    }
+
+    /// Compute image embeddings entirely on GPU (no CPU offload).
+    ///
+    /// Loads SAM on CUDA, runs global + patches, drops SAM.
+    /// Loads CLIP on CUDA, runs global + patches (with cached SAM outputs), drops CLIP.
+    /// All processing (projector, formatting, assembly) stays on CUDA.
+    /// Uses SequentialVramSwap internally to load one vision model at a time.
+    fn compute_image_embeddings_full_gpu(
+        &self,
+        inputs: &[Option<VisionInput<'_>>],
+        chunk_size: usize,
+    ) -> Result<Vec<Tensor>> {
+        // LM stays on CUDA (~950MB). Global SAM+CLIP on CPU. Patches on CUDA in
+        // mini-batches via SequentialVramSwap (~200MB SAM + ~200MB CLIP temporary).
+        // Peak VRAM: ~1.6GB — fits on 4GB without eviction.
+        let compute_dtype = self.dtype;
+        let projector = self.projector_for_dtype(compute_dtype);
+        let device = self.device();
+        let hidden_size = projector.hidden_size();
+        let newline = projector.image_newline_token(compute_dtype, device)?;
+        let view_sep = projector.view_separator_token(compute_dtype, device)?;
+        let cpu_vision = self.vision_modules().context("vision modules missing")?;
+        let wp = self.weights_path.clone();
+        let cfg = self.cfg.clone();
+        let dt = self.dtype;
+
+        let mut results = Vec::with_capacity(inputs.len());
+
+        for mb_input in inputs {
+            let result = match mb_input {
+                Some(vi) => {
+                    // === Global view: SAM + CLIP on CPU (avoids large attention score on CUDA) ===
+                    let global_cpu = prepare_image_tensor_for_device(
+                        vi.global, &Device::Cpu, compute_dtype,
+                    )?;
+                    let sg = cpu_vision
+                        .sam
+                        .forward(&global_cpu)
+                        .context("global SAM forward (CPU)")?;
+                    let cg = cpu_vision
+                        .clip
+                        .forward(&global_cpu, Some(&sg))
+                        .context("global CLIP forward (CPU)")?;
+                    let (_, clip_seq, _) = cg.shape().dims3()?;
+                    let ct = cg
+                        .narrow(D::Minus2, 1, clip_seq - 1)?
+                        .contiguous()?;
+                    let (_, sc, sh, sw) = sg.shape().dims4()?;
+                    let st = sg
+                        .reshape((1, sc, sh * sw))?
+                        .transpose(1, 2)?
+                        .contiguous()?;
+                    let combined_global = Tensor::cat(&[ct, st], D::Minus1)
+                        .context("cat global combined")?;
+                    // Move to CUDA for projector
+                    let combined_global_gpu = combined_global.to_device(device)?;
+
+                    // === Patches: SAM + CLIP on CUDA via sequential swap ===
+                    let local_tokens_opt: Option<Tensor> = vi
+                        .patches
+                        .as_ref()
+                        .map(|patches| -> Result<Tensor> {
+                            let patches_cpu = prepare_image_tensor_for_device(
+                                patches, &Device::Cpu, compute_dtype,
+                            )?;
+                            let (batch, _, _, _) = patches_cpu.shape().dims4()?;
+                            let swap =
+                                SequentialVramSwap::new(&wp, cfg.clone(), dt, device);
+                            let num = (batch + chunk_size - 1) / chunk_size;
+                            let pcs = patches_cpu.chunk(num, 0)?;
+
+                            let sam_cuda = swap
+                                .load_sam_on_cuda()
+                                .context("load SAM on CUDA for patches")?;
+                            let mut sam_cpu = Vec::with_capacity(num);
+                            for c in &pcs {
+                                let cc = c.to_device(device)?;
+                                let o = sam_cuda
+                                    .forward(&cc)
+                                    .context("patch SAM forward (CUDA)")?;
+                                sam_cpu.push(o.to_device(&Device::Cpu)?.contiguous()?);
+                            }
+                            drop(sam_cuda);
+
+                            let clip_cuda = swap
+                                .load_clip_on_cuda()
+                                .context("load CLIP on CUDA for patches")?;
+                            let mut all_proj = Vec::with_capacity(num);
+                            for (pc, sc) in pcs.iter().zip(sam_cpu.iter()) {
+                                let cc = pc.to_device(device)?;
+                                let st_cuda = sc.to_device(device)?;
+                                let co = clip_cuda
+                                    .forward(&cc, Some(&st_cuda))
+                                    .context("patch CLIP forward (CUDA)")?;
+                                let (_, cseq, _) = co.shape().dims3()?;
+                                let pct = co
+                                    .narrow(D::Minus2, 1, cseq - 1)?
+                                    .contiguous()?;
+                                let (_, scn, sh, sw) = st_cuda.shape().dims4()?;
+                                let pst = st_cuda
+                                    .reshape((cc.dims()[0], scn, sh * sw))?
+                                    .transpose(1, 2)?
+                                    .contiguous()?;
+                                let cb = Tensor::cat(&[pct, pst], D::Minus1)?;
+                                all_proj.push(projector.project(&cb)?);
+                            }
+                            drop(clip_cuda);
+
+                            let refs: Vec<&Tensor> = all_proj.iter().collect();
+                            let proj = Tensor::cat(&refs, 0)?;
+                            let cs = vi.crop_shape.unwrap_or((1, batch));
+                            let (wc, hc) = cs;
+                            let (_, seq, hd) = proj.shape().dims3()?;
+                            let side = (seq as f64).sqrt() as usize;
+                            let grid = proj
+                                .reshape((hc, wc, side, side, hd))?
+                                .permute((0, 2, 1, 3, 4))?
+                                .reshape((hc * side, wc * side, hd))?
+                                .contiguous()?;
+                            let (rows, cols, _) = grid.shape().dims3()?;
+                            let nle = newline
+                                .reshape((1, 1, hd))?
+                                .expand((rows, 1, hd))?
+                                .contiguous()?;
+                            Ok(Tensor::cat(&[grid, nle], 1)?
+                                .reshape((rows * (cols + 1), hd))?)
+                        })
+                        .transpose()?;
+
+                    // === Project global + format on CUDA ===
+                    let projected_global = projector.project(&combined_global_gpu)?;
+                    let nl = cast_dtype(&newline, projected_global.dtype(), "nl")?;
+                    let (_, gseq, ghd) = projected_global.shape().dims3()?;
+                    let side = (gseq as f64).sqrt() as usize;
+                    let grid = projected_global
+                        .get(0)?
+                        .reshape((side, side, ghd))?
+                        .contiguous()?;
+                    let nle = nl
+                        .reshape((1, 1, ghd))?
+                        .expand((side, 1, ghd))?
+                        .contiguous()?;
+                    let global_tokens = Tensor::cat(&[grid, nle], 1)?
+                        .reshape((side * (side + 1), ghd))?;
+
+                    // === Assemble ===
+                    let target_dtype = global_tokens.dtype();
+                    let mut segments = Vec::new();
+                    if let Some(lt) = local_tokens_opt {
+                        segments.push(
+                            cast_dtype_owned(lt, target_dtype, "local dtype cast")?,
+                        );
+                    }
+                    segments.push(global_tokens);
+                    let vst = view_sep.reshape((1, hidden_size))?.contiguous()?;
+                    segments.push(cast_dtype_owned(vst, target_dtype, "sep cast")?);
+                    Tensor::cat(&segments, 0)?
+                }
+                None => Tensor::zeros((0, hidden_size), compute_dtype, device)?,
+            };
+            results.push(result);
+        }
+
+        info!("Image embeddings computed (global on CPU, patches on CUDA via swap)");
+        Ok(results)
+    }
+
     pub fn compute_vision_projection(
         &self,
         input: &VisionInput<'_>,
@@ -1690,7 +2266,7 @@ impl DeepseekOcrModel {
                 let dtype = low_precision_compute_dtype(self.dtype);
                 let vision = select_f32(dtype, vision.as_ref(), self.vision_modules_f32());
                 let projector = self.projector_for_dtype(dtype);
-                VisionContext::new_with_dtype(self, vision, projector, dtype)
+                VisionContext::new_with_dtype(self, &vision.sam, &vision.clip, projector, dtype)
                     .process_input_full(input)
             }
             VisionBackend::Ocr2(_) => {
@@ -1986,9 +2562,10 @@ impl DeepseekOcrModel {
             let token_index = usize::try_from(current)
                 .context("token id out of range while preparing decode embedding")?;
             let mut decode_inputs = self
-                .language
+                .language_model()
                 .token_embedding_for_id(token_index)
                 .context("failed to gather embedding for decode token")?
+                .to_device(self.device())?
                 .unsqueeze(0)?
                 .unsqueeze(0)?;
             decode_inputs = cast_dtype_owned(
@@ -2384,7 +2961,7 @@ impl OcrEngine for DeepseekOcrModel {
             vision.crop_mode,
         )
         .with_context(|| "vision input failed")?;
-        let embeddings = compute_image_embeddings(self, &owned_inputs)
+        let embeddings = compute_image_embeddings(self, &owned_inputs, vision)
             .with_context(|| "image embedding failed")?;
         let (input_ids_vec, mask_vec) = build_prompt_tokens(
             tokenizer,
@@ -2434,6 +3011,8 @@ impl OcrEngine for DeepseekOcrModel {
             .into_iter()
             .next()
             .unwrap_or_default();
+        let gen_first: Vec<i64> = generated_tokens.iter().take(20).copied().collect();
+        trace!("decode: first 20 generated tokens = {gen_first:?}");
         let decoded = tokenizer
             .decode(
                 &generated_tokens
@@ -2443,6 +3022,7 @@ impl OcrEngine for DeepseekOcrModel {
                 true,
             )
             .unwrap_or_default();
+        trace!("decode: decoded text length = {} bytes", decoded.len());
         let normalized = normalize_text(&decoded);
 
         Ok(DecodeOutcome {
@@ -2494,6 +3074,7 @@ fn prepare_vision_inputs(
 fn compute_image_embeddings(
     model: &DeepseekOcrModel,
     owned_inputs: &[OwnedVisionInput],
+    vision: VisionSettings,
 ) -> Result<Vec<Tensor>> {
     let timer = Timer::new("vision.compute_embeddings");
     if owned_inputs.is_empty() {
@@ -2507,7 +3088,7 @@ fn compute_image_embeddings(
         .map(|owned| Some(owned.as_ref()))
         .collect();
     trace!("Computing image embeddings for {} image(s)...", refs.len());
-    let outputs = model.compute_image_embeddings(&refs);
+    let outputs = model.compute_image_embeddings_impl(&refs, &vision);
     match &outputs {
         Ok(values) => {
             let tokens_total: u64 = values
diff --git a/crates/infer-deepseek/src/model/swap.rs b/crates/infer-deepseek/src/model/swap.rs
new file mode 100644
index 0000000..753c205
--- /dev/null
+++ b/crates/infer-deepseek/src/model/swap.rs
@@ -0,0 +1,132 @@
+use std::sync::Arc;
+use std::path::Path;
+use std::time::Instant;
+
+use anyhow::{Context, Result};
+use candle_core::{DType, Device};
+use candle_nn::VarBuilder;
+use tracing::info;
+
+use crate::config::DeepseekOcrConfig;
+use crate::vision::{ClipVisionModel, SamBackbone};
+
+/// Detects total VRAM in bytes for the given CUDA device.
+/// Tries nvidia-smi first, then falls back to `DEEPSEEK_OCR_VRAM_MB` env var.
+pub(crate) fn get_vram_bytes(device: &Device) -> Option<u64> {
+    if !device.is_cuda() {
+        return None;
+    }
+    if let Ok(output) = std::process::Command::new("nvidia-smi")
+        .args(["--query-gpu=memory.total", "--format=csv,noheader,nounits"])
+        .output()
+    {
+        if let Ok(s) = String::from_utf8(output.stdout) {
+            if let Ok(mb) = s.trim().parse::<u64>() {
+                return Some(mb * 1024 * 1024);
+            }
+        }
+    }
+    if let Ok(vram_mb) = std::env::var("DEEPSEEK_OCR_VRAM_MB") {
+        if let Ok(mb) = vram_mb.parse::<u64>() {
+            return Some(mb * 1024 * 1024);
+        }
+    }
+    None
+}
+
+/// Whether to use sequential VRAM swap (low-VRAM devices < 6GB).
+pub(crate) fn should_use_vram_swap(device: &Device) -> bool {
+    match get_vram_bytes(device) {
+        Some(bytes) => bytes < 6 * 1024 * 1024 * 1024,
+        None => false,
+    }
+}
+
+/// Manages sequential VRAM swap for low-VRAM CUDA devices.
+///
+/// Never holds more than one heavy model in VRAM at a time:
+///   Phase 1: SAM on CUDA (~1.26 GB + activations ≈ 1.6 GB peak)
+///   Phase 2: CLIP on CUDA (~0.86 GB + activations ≈ 1.1 GB peak)
+///
+/// The LM (Q4K, ~950 MB) stays on CUDA permanently because QMatMul
+/// (quantized matrix multiply) does not support device transfer.
+///
+/// High-VRAM devices (>6 GB) skip this entirely and run normally.
+pub(crate) struct SequentialVramSwap {
+    weights_path: std::path::PathBuf,
+    cfg: Arc<DeepseekOcrConfig>,
+    dtype: DType,
+    cuda_device: Device,
+}
+
+impl SequentialVramSwap {
+    pub(crate) fn new(
+        weights_path: &Path,
+        cfg: Arc<DeepseekOcrConfig>,
+        dtype: DType,
+        cuda_device: &Device,
+    ) -> Self {
+        info!("SequentialVramSwap initialized — one vision model at a time on CUDA");
+        Self {
+            weights_path: weights_path.to_path_buf(),
+            cfg,
+            dtype,
+            cuda_device: cuda_device.clone(),
+        }
+    }
+
+    /// Load SAM backbone on CUDA from the safetensor file.
+    /// Caller MUST `drop()` the result before calling `load_clip_on_cuda()`
+    /// or running LM generation.
+    pub(crate) fn load_sam_on_cuda(&self) -> Result<SamBackbone> {
+        let t0 = Instant::now();
+        let vb = unsafe {
+            VarBuilder::from_mmaped_safetensors(
+                &[self.weights_path.as_path()],
+                self.dtype,
+                &self.cuda_device,
+            )
+        }
+        .with_context(|| {
+            format!(
+                "failed to mmap SAM weights on CUDA from {}",
+                self.weights_path.display()
+            )
+        })?;
+        let sam = SamBackbone::new(self.cfg.as_ref(), &vb.pp("model").pp("sam_model"))
+            .context("failed to load SAM backbone on CUDA")?;
+        info!(
+            elapsed = %format!("{:.2}s", t0.elapsed().as_secs_f32()),
+            "SAM loaded on CUDA"
+        );
+        Ok(sam)
+    }
+
+    /// Load CLIP vision model on CUDA from the safetensor file.
+    /// Caller MUST `drop()` the result before calling `load_sam_on_cuda()`
+    /// or running LM generation.
+    pub(crate) fn load_clip_on_cuda(&self) -> Result<ClipVisionModel> {
+        let t0 = Instant::now();
+        let vb = unsafe {
+            VarBuilder::from_mmaped_safetensors(
+                &[self.weights_path.as_path()],
+                self.dtype,
+                &self.cuda_device,
+            )
+        }
+        .with_context(|| {
+            format!(
+                "failed to mmap CLIP weights on CUDA from {}",
+                self.weights_path.display()
+            )
+        })?;
+        let clip = ClipVisionModel::load(self.cfg.as_ref(), &vb.pp("model").pp("vision_model"))
+            .context("failed to load CLIP vision model on CUDA")?;
+        info!(
+            elapsed = %format!("{:.2}s", t0.elapsed().as_secs_f32()),
+            "CLIP loaded on CUDA"
+        );
+        Ok(clip)
+    }
+
+}
diff --git a/crates/infer-deepseek/src/transformer/model.rs b/crates/infer-deepseek/src/transformer/model.rs
index 1c6ecc1..949006a 100644
--- a/crates/infer-deepseek/src/transformer/model.rs
+++ b/crates/infer-deepseek/src/transformer/model.rs
@@ -1,7 +1,7 @@
 use std::sync::Arc;
 
 use anyhow::{Context, Result, ensure};
-use candle_core::{DType, IndexOp, Tensor, quantized::QMatMul};
+use candle_core::{DType, Device, IndexOp, Tensor, quantized::QMatMul};
 use candle_nn::ops::{rms_norm, rms_norm_slow};
 use deepseek_ocr_core::tensor::gather_token_embeddings;
 
@@ -12,7 +12,7 @@ use crate::{
     transformer::{
         cache::{DynamicCache, PromptCacheGuard},
         decoder::TransformerDecoder,
-        weights::{DeepseekLanguageModelWeights, TransformerWeights},
+        weights::{DeepseekLanguageModelWeights, RmsNormWeights, TransformerWeights, qmatmul_to_device},
     },
 };
 
@@ -112,6 +112,13 @@ impl DeepseekLanguageModel {
         self.decoder.flash_attention_enabled()
     }
 
+    /// Move the token embedding tensor to the given device.
+    /// Useful on CUDA to keep the embedding on CPU and save VRAM for activations.
+    pub fn move_token_embedding_to(&mut self, device: &Device) -> Result<()> {
+        self.token_embedding = self.token_embedding.to_device(device)?;
+        Ok(())
+    }
+
     /// Lookup token embeddings for the provided input ids.
     pub fn embed_tokens(&self, input_ids: &Tensor) -> Result<Tensor> {
         let ids = if input_ids.dtype() == DType::I64 {
@@ -276,4 +283,41 @@ impl DeepseekLanguageModel {
             aux_loss: decoder_out.aux_loss,
         })
     }
+
+    /// Reconstruct this language model with all tensors moved to `device`.
+    /// Creates a fresh decoder (with empty RoPE cache) on the target device.
+    pub fn to_device(&self, device: &Device) -> Result<Self> {
+        let weights = DeepseekLanguageModelWeights {
+            token_embedding: self.token_embedding.to_device(device)?,
+            transformer: self.transformer_weights.to_device(device)?,
+            final_layernorm: RmsNormWeights {
+                weight: self.final_layernorm.to_device(device)?,
+            },
+            lm_head_weight: self
+                .lm_head_weight
+                .as_ref()
+                .map(|w| w.to_device(device))
+                .transpose()?,
+            lm_head_q: self
+                .lm_head_q
+                .as_ref()
+                .map(|qm| Ok::<_, anyhow::Error>(Arc::new(qmatmul_to_device(qm, device)?)))
+                .transpose()?,
+            lm_out_dim: self.lm_out_dim,
+            lm_in_dim: self.lm_in_dim,
+            lm_head_label: self.lm_head_label.clone(),
+        };
+        let mut model = Self::from_weights(self.cfg.clone(), weights);
+        model.final_layernorm_f32 = self
+            .final_layernorm_f32
+            .as_ref()
+            .map(|t| t.to_device(device))
+            .transpose()?;
+        model.lm_head_weight_f32 = self
+            .lm_head_weight_f32
+            .as_ref()
+            .map(|t| t.to_device(device))
+            .transpose()?;
+        Ok(model)
+    }
 }
diff --git a/crates/infer-deepseek/src/transformer/weights.rs b/crates/infer-deepseek/src/transformer/weights.rs
index c097c81..d371c6c 100644
--- a/crates/infer-deepseek/src/transformer/weights.rs
+++ b/crates/infer-deepseek/src/transformer/weights.rs
@@ -3,6 +3,11 @@ use std::{
     sync::Arc,
 };
 
+use anyhow::{Context, Result, ensure};
+use candle_core::{DType, Device, Tensor, quantized::{QMatMul, QStorage, QTensor}};
+use candle_nn::VarBuilder;
+use tracing::trace;
+
 use crate::{
     config::DeepseekV2Config,
     quant_snapshot::{
@@ -12,10 +17,6 @@ use crate::{
         LinearLayerGroup, QuantModule, QuantizationOutcome, QuantizationState, backend_label,
     },
 };
-use anyhow::{Context, Result, ensure};
-use candle_core::{DType, Tensor, quantized::QMatMul};
-use candle_nn::VarBuilder;
-use tracing::trace;
 
 /// Fully connected layer weights captured directly from safetensors via [`VarBuilder`].
 #[derive(Clone)]
@@ -64,7 +65,7 @@ impl LinearWeights {
             .with_context(|| format!("missing linear weight `{label}`"))?
             .contiguous()?;
         let mut weight = Some(weight_init.clone());
-        let weight_f32 = if matches!(weight_init.dtype(), DType::F16 | DType::BF16) {
+        let mut weight_f32 = if matches!(weight_init.dtype(), DType::F16 | DType::BF16) {
             Some(weight_init.to_dtype(DType::F32)?.contiguous()?)
         } else {
             None
@@ -101,6 +102,7 @@ impl LinearWeights {
                     bias_tensor = bias;
                     qmatmul = Some(qm);
                     weight = None;
+                    weight_f32 = None;
                 }
                 SnapshotLinear::Float {
                     weight: snapshot_weight,
@@ -121,6 +123,7 @@ impl LinearWeights {
                     quant.record_attempt(module, QuantizationOutcome::Fallback);
                     bias_tensor = bias;
                     weight = Some(snapshot_weight);
+                    weight_f32 = None;
                 }
             }
         }
@@ -637,3 +640,163 @@ pub(crate) fn qualified_name(vb: &VarBuilder, tensor: &str) -> String {
 }
 
 // Runtime quantization path removed: no `maybe_quantize_linear` fallback.
+
+// ---------------------------------------------------------------------------
+// Device transfer: used by the VRAM-swap path to move LM weights between CPU
+// and GPU while preserving quantization (QMatMul::QTensor reconstructed via
+// raw byte transfer).
+// ---------------------------------------------------------------------------
+
+/// Copy a [`QMatMul`] to another device, preserving its quantized form when
+/// possible (the `QTensor` variant gets reconstructed from raw bytes).
+pub(crate) fn qmatmul_to_device(qm: &QMatMul, device: &Device) -> Result<QMatMul> {
+    match qm {
+        QMatMul::QTensor(qt) => {
+            let data = qt.data()?;
+            let dtype = qt.dtype();
+            let shape = qt.shape().clone();
+            let storage = QStorage::from_data(data, device, dtype)?;
+            let qt_new = QTensor::new(storage, shape)?;
+            Ok(QMatMul::QTensor(Arc::new(qt_new)))
+        }
+        QMatMul::Tensor(t) => Ok(QMatMul::Tensor(t.to_device(device)?)),
+        QMatMul::TensorF16(t) => Ok(QMatMul::TensorF16(t.to_device(device)?)),
+    }
+}
+
+impl LinearWeights {
+    pub fn to_device(&self, device: &Device) -> Result<Self> {
+        Ok(Self {
+            weight: self
+                .weight
+                .as_ref()
+                .map(|w| w.to_device(device))
+                .transpose()?,
+            weight_f32: self
+                .weight_f32
+                .as_ref()
+                .map(|w| w.to_device(device))
+                .transpose()?,
+            bias: self
+                .bias
+                .as_ref()
+                .map(|b| b.to_device(device))
+                .transpose()?,
+            qmatmul: self
+                .qmatmul
+                .as_ref()
+                .map(|qm| Ok::<_, anyhow::Error>(Arc::new(qmatmul_to_device(qm, device)?)))
+                .transpose()?,
+            out_dim: self.out_dim,
+            in_dim: self.in_dim,
+            label: self.label.clone(),
+        })
+    }
+}
+
+impl RmsNormWeights {
+    pub fn to_device(&self, device: &Device) -> Result<Self> {
+        Ok(Self {
+            weight: self.weight.to_device(device)?,
+        })
+    }
+}
+
+impl AttentionWeights {
+    pub fn to_device(&self, device: &Device) -> Result<Self> {
+        Ok(Self {
+            q_proj: self.q_proj.to_device(device)?,
+            k_proj: self.k_proj.to_device(device)?,
+            v_proj: self.v_proj.to_device(device)?,
+            o_proj: self.o_proj.to_device(device)?,
+        })
+    }
+}
+
+impl DenseMlpWeights {
+    pub fn to_device(&self, device: &Device) -> Result<Self> {
+        Ok(Self {
+            gate_proj: self.gate_proj.to_device(device)?,
+            up_proj: self.up_proj.to_device(device)?,
+            down_proj: self.down_proj.to_device(device)?,
+        })
+    }
+}
+
+impl MoeWeights {
+    pub fn to_device(&self, device: &Device) -> Result<Self> {
+        Ok(Self {
+            gate_weight: self.gate_weight.to_device(device)?,
+            experts: self
+                .experts
+                .iter()
+                .map(|e| e.to_device(device))
+                .collect::<Result<Vec<_>>>()?,
+            shared_experts: self
+                .shared_experts
+                .as_ref()
+                .map(|s| s.to_device(device))
+                .transpose()?,
+            aux_bias: self
+                .aux_bias
+                .as_ref()
+                .map(|b| b.to_device(device))
+                .transpose()?,
+        })
+    }
+}
+
+impl MlpWeights {
+    pub fn to_device(&self, device: &Device) -> Result<Self> {
+        match self {
+            Self::Dense(d) => Ok(Self::Dense(d.to_device(device)?)),
+            Self::Moe(m) => Ok(Self::Moe(m.to_device(device)?)),
+        }
+    }
+}
+
+impl TransformerBlockWeights {
+    pub fn to_device(&self, device: &Device) -> Result<Self> {
+        Ok(Self {
+            attention: self.attention.to_device(device)?,
+            mlp: self.mlp.to_device(device)?,
+            input_layernorm: self.input_layernorm.to_device(device)?,
+            post_attention_layernorm: self.post_attention_layernorm.to_device(device)?,
+        })
+    }
+}
+
+impl TransformerWeights {
+    pub fn to_device(&self, device: &Device) -> Result<Self> {
+        Ok(Self {
+            layers: self
+                .layers
+                .iter()
+                .map(|l| l.to_device(device))
+                .collect::<Result<Vec<_>>>()?,
+        })
+    }
+}
+
+impl DeepseekLanguageModelWeights {
+    pub fn to_device(&self, device: &Device) -> Result<Self> {
+        Ok(Self {
+            token_embedding: self.token_embedding.to_device(device)?,
+            transformer: self.transformer.to_device(device)?,
+            final_layernorm: self.final_layernorm.to_device(device)?,
+            lm_head_weight: self
+                .lm_head_weight
+                .as_ref()
+                .map(|w| w.to_device(device))
+                .transpose()?,
+            lm_head_q: self
+                .lm_head_q
+                .as_ref()
+                .map(|qm| Ok::<_, anyhow::Error>(Arc::new(qmatmul_to_device(qm, device)?)))
+                .transpose()?,
+            lm_out_dim: self.lm_out_dim,
+            lm_in_dim: self.lm_in_dim,
+            lm_head_label: self.lm_head_label.clone(),
+        })
+    }
+}
diff --git a/crates/infer-deepseek/tests/long_generation_baseline.rs b/crates/infer-deepseek/tests/long_generation_baseline.rs
index c111ed3..1e21f5d 100644
--- a/crates/infer-deepseek/tests/long_generation_baseline.rs
+++ b/crates/infer-deepseek/tests/long_generation_baseline.rs
@@ -253,6 +253,7 @@ fn run_one_baseline(baseline_dir: &Path) -> Result<()> {
             base_size,
             image_size,
             crop_mode,
+            ..Default::default()
         };
 
         let outcome = model.decode(