[Quant] Keep the flashinfer_deepgemm FP8 GEMM to 1 <= M < 32 (#32843)

Co-authored-by: Claude Opus 5 <noreply@anthropic.com>

python/sglang/srt/layers/quantization/fp8_utils.py

@@ -751,7 +751,27 @@ def flashinfer_deepgemm_w8a8_block_fp8_linear_with_fallback(
     # fp8_blockscale_gemm_sm90 requires: N % 64 == 0, K % 128 == 0
     shape_supported = weight.shape[0] % 64 == 0 and weight.shape[1] % 128 == 0
 
-    if not (shape_supported and dtype_supported):
+    # Keep this backend to 1 <= M < 32, mirroring vLLM's
+    # FlashInferFp8DeepGEMMDynamicBlockScaledKernel. fp8_blockscale_gemm_sm90 is
+    # one entry point over two kernels and only the M < 32 swapAB half is worth
+    # taking:
+    #   M >= 32 picks the non-swapAB kernel, which is slower than DeepGEMM (worst
+    #     just above the threshold) and, on some checkpoints, less accurate.
+    #   M == 0 hard-fails inside the kernel ("Check failed: (input_ptr !=
+    #     nullptr)"). Empty batches are a normal steady-state input, not an edge
+    #     case: DP attention hands an idle rank a zero-token forward so the
+    #     collectives stay in sync (ScheduleBatch.prepare_for_idle).
+    # Same shape of guard as the gfx95 CK M bound below.
+    m_supported = 1 <= input.view(-1, input.shape[-1]).shape[0] < 32
+
+    if not m_supported and deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM:
+        # DeepGEMM covers both ends and falls back to triton on its own for
+        # shapes it cannot serve.
+        return deepgemm_w8a8_block_fp8_linear_with_fallback(
+            input, weight, block_size, weight_scale, input_scale, bias
+        )
+
+    if not (shape_supported and dtype_supported and m_supported):
         if weight_scale.dtype == torch.int32:
             weight_scale = _unpack_ue8m0_scale_for_triton(
                 weight_scale, weight.shape, block_size