tune Qwen3-VL-4B prefill unified-attention on gfx1150

vllm/model_executor/kernels/linear/mixed_precision/hybrid_w4a16.py

@@ -23,7 +23,7 @@
     permute_param_layout_,
 )
 from vllm.platforms import current_platform
-from vllm.platforms.rocm import on_gfx1x, on_gfx1103, on_gfx1151
+from vllm.platforms.rocm import on_gfx1x, on_gfx1103, on_gfx1150, on_gfx1151
 from vllm.scalar_type import scalar_types
 from vllm.triton_utils import tl, triton
 from vllm.utils.torch_utils import direct_register_custom_op
@@ -275,6 +275,16 @@ def _select_skinny_gfx11_config(
 ) -> tuple[int, int, int, int]:
     """Return (BLOCK_M, BLOCK_N, BLOCK_K, num_warps) for the gfx11 skinny GEMM."""
     if dtype == torch.float16:
+        # Profile-guided default for Qwen3-VL-like multimodal prefill.
+        qwen3_prefill_shapes = {
+            (19456, 2560),  # gate_up_proj-like
+            (2560, 9728),  # down_proj-like
+            (6144, 2560),  # qkv_proj-like
+            (2560, 4096),  # o_proj-like
+        }
+        if on_gfx1150() and 576 <= M <= 832 and (N, K) in qwen3_prefill_shapes:
+            return 64, 256, 64, 8
+
         # Packed-dequant path (fp16 on gfx1x). Cold-optimal tiers from a broad
         # rotating-buffer cudagraph sweep over the catalog (weights row-padded
         # exactly as production via pack_skinny_int4, so the K%4096 cliff is

vllm/platforms/rocm.py

@@ -193,6 +193,7 @@ def _get_gcn_arch() -> str:
 _ON_GFX11 = "gfx11" in _GCN_ARCH
 _ON_GFX1100 = "gfx1100" in _GCN_ARCH
 _ON_GFX1103 = "gfx1103" in _GCN_ARCH
+_ON_GFX1150 = "gfx1150" in _GCN_ARCH
 _ON_GFX1151 = "gfx1151" in _GCN_ARCH
 _ON_GFX12X = any(arch in _GCN_ARCH for arch in ["gfx12"])
 _ON_MI3XX = any(arch in _GCN_ARCH for arch in ["gfx942", "gfx950"])
@@ -288,6 +289,8 @@ def on_gfx1100() -> bool:
 def on_gfx1103() -> bool:
     return _ON_GFX1103
 
+def on_gfx1150() -> bool:
+    return _ON_GFX1150
 
 def on_gfx1151() -> bool:
     return _ON_GFX1151

vllm/v1/attention/ops/triton_unified_attention.py

@@ -750,6 +750,25 @@
     return sliding_window == 1024 and head_size in (128, 256)
 
 
+def _is_qwen3_vl_4b_prefill_signature(
+    *,
+    head_size: int,
+    num_query_heads: int,
+    num_kv_heads: int,
+    max_seqlen_q: int,
+) -> bool:
+    """Detect Qwen3-VL-4B text prefill attention shape.
+
+    This signature is intentionally strict to avoid affecting unrelated models.
+    """
+    return (
+        head_size == 128
+        and num_query_heads == 32
+        and num_kv_heads == 8
+        and max_seqlen_q >= 256
+    )
+
+
 def _get_tile_size(
     head_size: int,
     sliding_window: int,
@@ -1088,15 +1107,35 @@
 
     grid: tuple[Any, ...]
     config = {}
+    tile_size_prefill = TILE_SIZE_PREFILL
 
     use_swapped_grid = not use_3d and current_platform.is_gfx1151()
 
+    # Optional tuning overrides for targeted profiling/experiments.
+    # These are intentionally no-op unless explicitly set.
+    if not use_3d:
+        # Qwen3-VL-4B prefill default on Navi/gfx11:
+        # BM64/T32/W4/S1/EU4 was the best end-to-end TTFT setting in local sweeps.
+        if current_platform.is_navi() and _is_qwen3_vl_4b_prefill_signature(
+            head_size=head_size,
+            num_query_heads=num_query_heads,
+            num_kv_heads=num_kv_heads,
+            max_seqlen_q=max_seqlen_q,
+        ):
+            BLOCK_M = 64
+            BLOCK_Q = BLOCK_M // num_queries_per_kv
+            total_num_q_blocks = q.shape[0] // BLOCK_Q + num_seqs
+            tile_size_prefill = 32
+            num_warps = 4
+            num_stages = 1
+            waves_per_eu = 4
+
     if not use_3d:
         if use_swapped_grid:
             grid = (num_kv_heads, total_num_q_blocks)
         else:
             grid = (total_num_q_blocks, num_kv_heads)
-        tile_size = TILE_SIZE_PREFILL
+        tile_size = tile_size_prefill
     else:
         tile_size = TILE_SIZE_DECODE