ssm_conv: channels-major input mode to drop the delta-net transpose

[roberteg16]

Summary

ggml_ssm_conv requires a time-major input [time, channels, seqs], while the delta-net projection produces channels-major [channels, time, seqs]. Feeding the op therefore required a physical transpose of the conv input on every layer, which is a significant kernel in prefill.

This PR adds an opt-in channels-major layout to ggml_ssm_conv (ggml_ssm_conv_channels_major): the input is [d_inner, d_conv-1+n_t, n_s] (channels contiguous) and the output layout is unchanged. build_conv_state keeps qkv_mixed channels-major, prepends the recurrent conv state along the time axis (concat dim 1), and calls the new entry point, so the transpose is no longer needed.

• Implemented on CPU and CUDA (short- and long-token kernels; coalesced smem load).
• Non-implementing backends fall back to CPU; existing (time-major) callers are bit-identical.
• Shared by the qwen35, qwen35moe, and qwen3next graphs.

The two commits are: (1) force the conv-state concat input contiguous; (2) add the channels-major ssm_conv mode and switch build_conv_state to it.

Measured impact — roofline (Qwen3.6-35B-A3B, prefill pp4096, gfx1151)

-ub 4096 -b 4096 (M=4096)

op	baseline	this PR
CONCAT	1166 ms	34 ms
CONT (transpose)	10 ms	10 ms
SSM_CONV	47 ms	43 ms
total prefill	3825 ms	2717 ms

−1108 ms (−29 %).

-ub 4096 -b 2048 (M=2048; n_ubatch is clamped to n_batch, which defaults to 2048)

op	baseline	this PR
CONCAT	283 ms	35 ms
total prefill	2786 ms	2590 ms

−196 ms (−7.0 %).

Measured impact — throughput A/B (llama-bench, Qwen3.6-35B-A3B, gfx1151, t/s, CUDA graphs enabled)

test	baseline	this PR	Δ
pp128	1061.2 ± 1.3	1052.4 ± 24.1	~0
pp512	1558.8 ± 16.0	1641.8 ± 28.6	+5.3 %
pp1024	1584.9 ± 16.0	1652.9 ± 53.8	+4.3 %
pp512 @ d512	1536.2 ± 3.8	1615.8 ± 19.6	+5.2 %
pp1024 @ d1024	1527.9 ± 45.5	1632.0 ± 24.2	+6.8 %
tg128 @ d128	54.77 ± 0.42	54.95 ± 0.43	~0
tg512 @ d512	54.17 ± 0.33	54.51 ± 0.20	~0
tg1024 @ d1024	54.32 ± 0.21	54.82 ± 0.03	~0

Baseline is the branch point 1847907d6.

Validation

Kernel numerics. test-backend-ops -o SSM_CONV runs the channels-major cases (short- and long-token paths, d_conv ∈ {3,4,9}, d_inner ∈ {1024,1536,2048}, single/multi-seq) against the CPU reference: 90/90 pass.

Output-preserving. Perplexity over a fixed ~500-token passage (-c 256, greedy) is bit-identical between baseline (1847907d6) and this PR on all three affected architectures:

arch	model (Q4_K_M)	baseline PPL	this PR PPL
qwen35	Qwen3.5-0.8B	16.2847	16.2847
qwen35moe	Qwen3.6-35B-A3B	3.4005	3.4005
qwen3next	Qwen3-Coder-Next-REAP-40B-A3B	26.5075	26.5075

Qwen3.5-35B-A3B (qwen35moe) additionally runs at PPL 2.18.

Notes / out of scope

• Non-CUDA backends only exercised via the CPU fallback path; not run on Metal/Vulkan/SYCL/etc.
• ggml-backend-meta.cpp tensor-parallel split handles the channels-major layout (d_inner is sx axis 0 / c axis 1). Not exercised in a tensor-parallel run here.

[Copilot]

Pull request overview

This PR introduces an opt-in channels-major input layout for ggml_ssm_conv and updates the Qwen3.5/3.6 hybrid (DeltaNet + MoE) graphs to use it, eliminating a costly transpose in prefill and improving large-ubatch performance.

Changes:

• Add GGML_SSM_CONV_LAYOUT_{TIME_MAJOR,CHANNELS_MAJOR}, a new entry point ggml_ssm_conv_channels_major(), and plumb layout through op params so CPU/CUDA can select the correct stride interpretation.
• Update DeltaNet conv-state construction in the Qwen graphs to keep [channels, time, seq] and call ggml_ssm_conv_channels_major() directly (dropping the DeltaNet-side transpose).
• Extend backend-op tests to exercise the new layout; gate non-CPU/CUDA backends to time-major until they implement channels-major.

Reviewed changes

Copilot reviewed 17 out of 17 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
tests/test-backend-ops.cpp	Adds coverage for the channels-major SSM_CONV mode in backend op tests.
src/models/qwen3next.cpp	Switches Qwen3Next to call ggml_ssm_conv_channels_major() for conv.
src/models/qwen35moe.cpp	Switches Qwen3.6 MoE graph to the channels-major SSM_CONV entry point.
src/models/qwen35.cpp	Switches Qwen3.5 graph to the channels-major SSM_CONV entry point.
src/models/delta-net-base.cpp	Reworks conv-state concat to be channels-major and removes the DeltaNet transpose.
ggml/src/ggml.c	Adds layout-aware SSM_CONV constructor + new channels-major API + layout getter.
ggml/src/ggml-webgpu/ggml-webgpu.cpp	Disables channels-major offload on WebGPU (time-major only).
ggml/src/ggml-vulkan/ggml-vulkan.cpp	Disables channels-major offload on Vulkan (time-major only).
ggml/src/ggml-sycl/ggml-sycl.cpp	Disables channels-major offload on SYCL (time-major only).
ggml/src/ggml-opencl/ggml-opencl.cpp	Disables channels-major offload on OpenCL (time-major only).
ggml/src/ggml-metal/ggml-metal-device.m	Disables channels-major offload on Metal (time-major only).
ggml/src/ggml-hexagon/ggml-hexagon.cpp	Disables channels-major offload on Hexagon (time-major only).
ggml/src/ggml-cuda/ssm-conv.cu	Implements channels-major-aware CUDA kernels for SSM_CONV (short/long token paths).
ggml/src/ggml-cuda/ggml-cuda.cu	Updates CUDA supports-op logic to interpret d_inner based on layout.
ggml/src/ggml-cpu/ops.cpp	Makes the CPU SSM_CONV implementation layout-aware via computed channel/time strides.
ggml/src/ggml-cann/ggml-cann.cpp	Disables channels-major offload on CANN (time-major only).
ggml/include/ggml.h	Exposes the new layout enum, channels-major entry point, and layout getter in the public API.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 18 out of 18 changed files in this pull request and generated 1 comment.

[Copilot]

Pull request overview

Copilot reviewed 18 out of 18 changed files in this pull request and generated no new comments.

[mgehre-amd]

Looks good