[AMD] Use tgemm.mm for MoEGate router gemm in deepseek_v2.py (#21657)

python/sglang/srt/layers/rocm_linear_utils.py

@@ -1,10 +1,7 @@
 import torch
 from aiter.ops.triton.fused_kv_cache import fused_qk_rope_cat_and_cache_mla
 from aiter.ops.triton.fused_qk_concat import fused_qk_rope_cat
-from aiter.ops.triton.gemm_a16w16 import gemm_a16w16
+from aiter.tuned_gemm import tgemm
-from aiter.ops.triton.gemm_a16w16_atomic import gemm_a16w16_atomic
-
-from sglang.srt.utils import BumpAllocator
 
 __all__ = ["fused_qk_rope_cat", "fused_qk_rope_cat_and_cache_mla"]
 
@@ -12,26 +9,9 @@ __all__ = ["fused_qk_rope_cat", "fused_qk_rope_cat_and_cache_mla"]
 def aiter_dsv3_router_gemm(
     hidden_states: torch.Tensor,
     weight: torch.Tensor,
-    gemm_output_zero_allocator: BumpAllocator = None,
 ):
-    M = hidden_states.shape[0]
+    """Use aiter tuned GEMM dispatcher (tgemm.mm) to automatically select the GEMM kernel."""
-    N = weight.shape[0]
+    return tgemm.mm(hidden_states, weight, otype=hidden_states.dtype)
-    y = None
-
-    if M <= 256:
-        # TODO (cagri): convert to bfloat16 as part of another kernel to save time
-        # for now it is also coupled with zero allocator.
-        if gemm_output_zero_allocator != None:
-            y = gemm_output_zero_allocator.allocate(M * N).view(M, N)
-        else:
-            y = torch.zeros((M, N), dtype=torch.float32, device=hidden_states.device)
-
-    if y is not None:
-        logits = gemm_a16w16_atomic(hidden_states, weight, y=y).to(hidden_states.dtype)
-    else:
-        logits = gemm_a16w16(hidden_states, weight)
-
-    return logits
 
 
 def get_dsv3_gemm_output_zero_allocator_size(

python/sglang/srt/models/deepseek_v2.py

@@ -153,9 +153,11 @@ from sglang.srt.utils import (
     use_intel_amx_backend,
 )
 
+if _use_aiter:
+    from sglang.srt.layers.rocm_linear_utils import aiter_dsv3_router_gemm
+
 if _use_aiter_gfx95:
     from sglang.srt.layers.rocm_linear_utils import (
-        aiter_dsv3_router_gemm,
         get_dsv3_gemm_output_zero_allocator_size,
     )
 
@@ -327,14 +329,8 @@ class MoEGate(nn.Module):
                 logits = dsv3_router_gemm(
                     hidden_states, self.weight, out_dtype=torch.float32
                 )
-            elif (
+            elif _use_aiter:
-                _use_aiter_gfx95
+                logits = aiter_dsv3_router_gemm(hidden_states, self.weight)
-                and hidden_states.shape[0] <= 256
-                and self.weight.shape[0] <= 256
-            ):
-                logits = aiter_dsv3_router_gemm(
-                    hidden_states, self.weight, gemm_output_zero_allocator
-                )
             else:
                 logits = F.linear(hidden_states, self.weight, None)