[ROCm] triton fp8 kernel (#27058)

Signed-off-by: Aleksandr Malyshev <maleksan@amd.com>
Co-authored-by: Aleksandr Malyshev <maleksan@amd.com>
Co-authored-by: Gregory Shtrasberg <156009573+gshtras@users.noreply.github.com>

vllm/model_executor/layers/quantization/utils/fp8_utils.py

@@ -69,30 +69,67 @@ def cutlass_scaled_mm(
 
 
 def rocm_aiter_gemm_w8a8_blockscale_impl(
-    A: torch.Tensor,
+    input_2d: torch.Tensor,
-    B: torch.Tensor,
+    weight: torch.Tensor,
-    As: torch.Tensor,
+    input_scale: torch.Tensor,
-    Bs: torch.Tensor,
+    weight_scale: torch.Tensor,
-    block_size: list[int],
+    group_size: int,
     output_dtype: torch.dtype = torch.float16,
 ) -> torch.Tensor:
-    import aiter as rocm_aiter
+    def is_aiter_triton_kernel_tuned(n, k):
+        return (n, k) in [
+            (1024, 8192),
+            (2112, 7168),
+            (3072, 1536),
+            (32768, 8192),
+            (4096, 7168),
+            (4608, 7168),
+            (512, 7168),
+            (7168, 2048),
+            (7168, 256),
+            (8192, 1024),
+            (8192, 32768),
+        ]
 
-    return rocm_aiter.gemm_a8w8_blockscale(A, B, As, Bs, dtype=output_dtype)
+    n, k = weight.shape
+    if input_scale is not None:
+        q_input = input_2d
+    elif not current_platform.is_fp8_fnuz() and is_aiter_triton_kernel_tuned(n, k):
+        from aiter.ops.triton.gemm_a8w8_blockscale import gemm_a8w8_blockscale
+
+        # MI350 case uses triton kernel
+        q_input, input_scale = per_token_group_quant_fp8(
+            input_2d,
+            group_size,
+            column_major_scales=False,
+            use_ue8m0=False,
+        )
+    else:
+        # MI300 uses tuned AITER ASM/C++ kernel
+        import aiter as rocm_aiter
+        from aiter import gemm_a8w8_blockscale, get_hip_quant
+
+        aiter_per1x128_quant = get_hip_quant(rocm_aiter.QuantType.per_1x128)
+        q_input, input_scale = aiter_per1x128_quant(
+            input_2d.contiguous(), quant_dtype=rocm_aiter.dtypes.fp8
+        )
+
+    return gemm_a8w8_blockscale(
+        q_input, weight, input_scale, weight_scale, dtype=output_dtype
+    )
 
 
 def rocm_aiter_gemm_w8a8_blockscale_fake(
-    A: torch.Tensor,
+    input_2d: torch.Tensor,
-    B: torch.Tensor,
+    weight: torch.Tensor,
-    As: torch.Tensor,
+    input_scale: torch.Tensor,
-    Bs: torch.Tensor,
+    weight_scale: torch.Tensor,
-    block_size: list[int],
+    group_size: int,
     output_dtype: torch.dtype = torch.float16,
 ) -> torch.Tensor:
-    m = A.shape[0]
+    m = input_2d.shape[0]
-    n = B.shape[0]
+    n = weight.shape[0]
-    Y = torch.empty(m, n, dtype=output_dtype, device=A.device)
+    return torch.empty(m, n, dtype=output_dtype, device=input_2d.device)
-    return Y
 
 
 if current_platform.is_rocm():
@@ -101,15 +138,6 @@ if current_platform.is_rocm():
         op_func=rocm_aiter_gemm_w8a8_blockscale_impl,
         fake_impl=rocm_aiter_gemm_w8a8_blockscale_fake,
     )
-    if (
-        envs.VLLM_ROCM_USE_AITER
-        and envs.VLLM_ROCM_USE_AITER_LINEAR
-        and current_platform.is_fp8_fnuz()
-    ):
-        import aiter as rocm_aiter
-        from aiter import get_hip_quant
-
-        aiter_per1x128_quant = get_hip_quant(rocm_aiter.QuantType.per_1x128)
 
 
 # TODO we should be able to change the type of block_size to GroupShape
@@ -293,7 +321,9 @@ class W8A8BlockFp8LinearOp:
         ):
             output = self._run_deepgemm(input_2d, weight, weight_scale)
         else:
-            output = self.w8a8_blockscale_op(input_2d, weight, weight_scale)
+            output = self.w8a8_blockscale_op(
+                input_2d, weight, weight_scale, input_scale
+            )
 
         if bias is not None:
             output = output + bias
@@ -322,7 +352,9 @@ class W8A8BlockFp8LinearOp:
         input_2d: torch.Tensor,
         weight: torch.Tensor,
         weight_scale: torch.Tensor,
+        input_scale: torch.Tensor | None = None,
     ) -> torch.Tensor:
+        assert input_scale is None
         assert self.input_quant_op is not None
         q_input, input_scale = self.input_quant_op(input_2d)
         if self.is_hopper:
@@ -350,17 +382,15 @@ class W8A8BlockFp8LinearOp:
         input_2d: torch.Tensor,
         weight: torch.Tensor,
         weight_scale: torch.Tensor,
+        input_scale: torch.Tensor | None = None,
     ) -> torch.Tensor:
         assert self.act_quant_group_shape == GroupShape(1, 128)
-        q_input, input_scale = aiter_per1x128_quant(
-            input_2d.contiguous(), quant_dtype=rocm_aiter.dtypes.fp8
-        )
         return torch.ops.vllm.rocm_aiter_gemm_w8a8_blockscale(
-            q_input,
+            input_2d,
             weight,
             input_scale,
             weight_scale,
-            list(self.weight_group_shape),
+            self.act_quant_group_shape.col,
             input_2d.dtype,
         )
 
@@ -369,7 +399,9 @@ class W8A8BlockFp8LinearOp:
         input_2d: torch.Tensor,
         weight: torch.Tensor,
         weight_scale: torch.Tensor,
+        input_scale: torch.Tensor | None = None,
     ) -> torch.Tensor:
+        assert input_scale is None
         assert self.input_quant_op is not None
         q_input, input_scale = self.input_quant_op(input_2d)
         return torch.ops.vllm.w8a8_triton_block_scaled_mm_func(
@@ -391,6 +423,7 @@ class W8A8BlockFp8LinearOp:
                 torch.Tensor,
                 torch.Tensor,
                 torch.Tensor,
+                torch.Tensor | None,
             ],
             torch.Tensor,
         ],
@@ -939,13 +972,11 @@ def requant_weight_ue8m0_inplace(
 
 
 def check_aiter_fp8_linear_support() -> bool:
-    """AITER is only supported on ROCm and only for FP8_FNUZ
+    """AITER is only supported on ROCm for MI3XX"""
-    and at the moment are MI300 series"""
     return (
         current_platform.is_rocm()
         and envs.VLLM_ROCM_USE_AITER
         and envs.VLLM_ROCM_USE_AITER_LINEAR
-        and current_platform.is_fp8_fnuz()
     )