[FEAT] [ROCm] Add AITER int8 scaled gemm kernel (#15433)

Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>

tests/quantization/test_compressed_tensors.py

@@ -20,6 +20,23 @@ from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
     sparse_cutlass_supported)
 from vllm.platforms import current_platform
 
+# AITER only supports per-channel-per-channel INT8 gemm
+# and per-tensor-per-tensor INT8 GEMM.
+# It does not support mix precision MM and mix quantization scheme.
+ROCM_AITER_SUPPORTED_INT8_MODEL = [
+    "neuralmagic/Llama-3.2-1B-quantized.w8a8",
+    "nm-testing/tinyllama-oneshot-w8a8-channel-dynamic-token-v2"
+]
+
+# TritonScaledMMLinearKernel only supports symmetric quantization.
+ROCM_TRITON_SCALED_MM_SUPPORTED_INT8_MODEL = [
+    "nm-testing/tinyllama-oneshot-w8w8-test-static-shape-change",
+    "nm-testing/tinyllama-oneshot-w8-channel-a8-tensor",
+    "neuralmagic/Llama-3.2-1B-quantized.w8a8",
+    "nm-testing/tinyllama-oneshot-w8a8-dynamic-token-v2",
+    "nm-testing/tinyllama-oneshot-w8a8-channel-dynamic-token-v2",
+]
+
 
 @pytest.fixture(scope="function", autouse=True)
 def use_v0_only(monkeypatch):
@@ -57,6 +74,11 @@ def use_v0_only(monkeypatch):
 )
 def test_compressed_tensors_w8a8_static_setup(vllm_runner, model_args):
     model_path, strategy, quant_type, shape_0, is_symmetric = model_args
+
+    if current_platform.is_rocm(
+    ) and model_path not in ROCM_TRITON_SCALED_MM_SUPPORTED_INT8_MODEL:
+        pytest.skip(f"Skip model {model_path} as it is not support on ROCm.")
+
     with vllm_runner(model_path, enforce_eager=True) as llm:
 
         def check_model(model):
@@ -123,6 +145,8 @@ def test_compressed_tensors_w8a8_static_setup(vllm_runner, model_args):
 )
 @pytest.mark.parametrize("max_tokens", [32])
 @pytest.mark.parametrize("num_logprobs", [10])
+@pytest.mark.parametrize(
+    "use_aiter", [True, False] if current_platform.is_rocm() else [False])
 def test_compressed_tensors_w8a8_logprobs(
     hf_runner,
     vllm_runner,
@@ -130,7 +154,21 @@ def test_compressed_tensors_w8a8_logprobs(
     model_path,
     max_tokens,
     num_logprobs,
+    use_aiter,
+    monkeypatch,
 ):
+
+    if current_platform.is_rocm(
+    ) and model_path not in ROCM_TRITON_SCALED_MM_SUPPORTED_INT8_MODEL:
+        pytest.skip(f"Skip model {model_path} as it is not support on ROCm.")
+
+    if use_aiter:
+        if model_path not in ROCM_AITER_SUPPORTED_INT8_MODEL:
+            pytest.skip(
+                f"Skip model {model_path} as it is not support by aiter.")
+        # this will enable VLLM_ROCM_USE_AITER_LINEAR
+        monkeypatch.setenv("VLLM_ROCM_USE_AITER", "1")
+
     dtype = "bfloat16"
 
     # skip language translation prompt for the static per tensor asym model
@@ -154,6 +192,9 @@ def test_compressed_tensors_w8a8_logprobs(
         name_1="vllm",
     )
 
+    if current_platform.is_rocm():
+        torch.cuda.synchronize()
+
 
 def test_compressed_tensors_no_enforce_eager(vllm_runner):
     model_path = "nm-testing/tinyllama-oneshot-w8w8-test-static-shape-change"
@@ -177,8 +218,27 @@ def test_compressed_tensors_no_enforce_eager(vllm_runner):
         ),
     ],
 )
-def test_compressed_tensors_w8a8_dynamic_per_token(vllm_runner, model_args):
+@pytest.mark.parametrize(
+    "use_aiter", [True, False] if current_platform.is_rocm() else [False])
+def test_compressed_tensors_w8a8_dynamic_per_token(
+    vllm_runner,
+    model_args,
+    use_aiter,
+    monkeypatch,
+):
     model_path, strategy = model_args
+
+    if current_platform.is_rocm(
+    ) and model_path not in ROCM_TRITON_SCALED_MM_SUPPORTED_INT8_MODEL:
+        pytest.skip(f"Skip model {model_path} as it is not support on ROCm.")
+
+    if use_aiter:
+        if model_path not in ROCM_AITER_SUPPORTED_INT8_MODEL:
+            pytest.skip(
+                f"Skip model {model_path} as it is not support by aiter.")
+        # this will enable VLLM_ROCM_USE_AITER_LINEAR
+        monkeypatch.setenv("VLLM_ROCM_USE_AITER", "1")
+
     with vllm_runner(model_path, dtype=torch.float16) as llm:
 
         def check_model(model):
@@ -207,6 +267,8 @@ def test_compressed_tensors_w8a8_dynamic_per_token(vllm_runner, model_args):
         ("nm-testing/tinyllama-oneshot-w8a16-per-channel", "channel", None, 4),
     ],
 )
+@pytest.mark.skipif(not current_platform.is_cuda(),
+                    reason="The tests are skipped on non-CUDA platform.")
 def test_compressed_tensors_wNa16(vllm_runner, wNa16_args):
     model, strategy, group, pack_factor = wNa16_args
     with vllm_runner(model) as llm:
@@ -231,6 +293,8 @@ def test_compressed_tensors_wNa16(vllm_runner, wNa16_args):
         assert output
 
 
+@pytest.mark.skipif(not current_platform.is_cuda(),
+                    reason="This test is skipped on non-CUDA platform.")
 def test_compressed_tensors_w4a16_marlin24(vllm_runner):
     model_path = "nm-testing/llama7b-one-shot-2_4-w4a16-marlin24-t"
     with vllm_runner(model_path) as llm:
@@ -271,7 +335,7 @@ def test_compressed_tensors_fp8(vllm_runner):
 
             if isinstance(qkv_proj.scheme, CompressedTensorsW8A8Fp8):
                 assert len(qkv_proj.input_scale.shape) == 0
-                assert qkv_proj.weight.dtype is torch.float8_e4m3fn
+                assert qkv_proj.weight.dtype is current_platform.fp8_dtype()
                 assert qkv_proj.weight_scale.dtype is torch.float32
                 assert len(qkv_proj.weight_scale.shape) == 0
 
@@ -281,6 +345,8 @@ def test_compressed_tensors_fp8(vllm_runner):
         assert output
 
 
+@pytest.mark.skipif(not current_platform.is_cuda(),
+                    reason="This test is skipped on non-CUDA platform.")
 def test_compressed_tensors_kv_cache(vllm_runner):
     model_path = "nm-testing/TinyLlama-1.1B-compressed-tensors-kv-cache-scheme"
     with vllm_runner(model_path, kv_cache_dtype="fp8") as llm:
@@ -309,7 +375,8 @@ def _test_2of4_quant_models(qkv_proj,
 
 
 @pytest.mark.skipif(
-    not current_platform.has_device_capability(90),
+    not current_platform.is_cuda()
+    or not current_platform.has_device_capability(90),
     reason="Sparse FP8 is not yet supported on this GPU type.",
 )
 @pytest.mark.parametrize(
@@ -356,7 +423,8 @@ def test_compressed_tensors_2of4_quant_fp8(vllm_runner, args_2of4):
 
 
 @pytest.mark.skipif(
-    not current_platform.has_device_capability(90),
+    not current_platform.is_cuda()
+    or not current_platform.has_device_capability(90),
     reason="Sparse FP8 is not yet supported on this GPU type.",
 )
 @pytest.mark.parametrize(

vllm/envs.py

@@ -75,6 +75,7 @@ if TYPE_CHECKING:
     VLLM_DISABLED_KERNELS: list[str] = []
     VLLM_USE_V1: bool = True
     VLLM_ROCM_USE_AITER: bool = False
+    VLLM_ROCM_USE_AITER_LINEAR: bool = True
     VLLM_ROCM_USE_AITER_MOE: bool = True
     VLLM_ROCM_USE_AITER_FP8_BLOCK_SCALED_MOE: bool = False
     VLLM_ROCM_USE_AITER_RMSNORM: bool = True
@@ -524,6 +525,13 @@ environment_variables: dict[str, Callable[[], Any]] = {
     lambda: (os.getenv("VLLM_ROCM_USE_AITER", "False").lower() in
              ("true", "1")),
 
+    # use aiter linear op if aiter ops are enabled
+    # The following list of related ops
+    # - scaled_mm (per-tensor / rowwise)
+    "VLLM_ROCM_USE_AITER_LINEAR":
+    lambda: (os.getenv("VLLM_ROCM_USE_AITER_LINEAR", "True").lower() in
+             ("true", "1")),
+
     # Whether to use aiter moe ops.
     # By default is enabled.
     "VLLM_ROCM_USE_AITER_MOE":

vllm/model_executor/layers/quantization/kernels/scaled_mm/__init__.py

@@ -3,6 +3,8 @@
 import os
 from typing import Dict, List, Optional, Type
 
+from vllm.model_executor.layers.quantization.kernels.scaled_mm.aiter import (
+    AiterScaledMMLinearKernel)
 from vllm.model_executor.layers.quantization.kernels.scaled_mm.cutlass import (
     CutlassScaledMMLinearKernel)
 from vllm.model_executor.layers.quantization.kernels.scaled_mm.ScaledMMLinearKernel import (  # noqa: E501
@@ -17,7 +19,7 @@ from vllm.platforms import PlatformEnum, current_platform
 _POSSIBLE_KERNELS: Dict[PlatformEnum, List[Type[ScaledMMLinearKernel]]] = {
     PlatformEnum.CPU: [CutlassScaledMMLinearKernel],
     PlatformEnum.CUDA: [CutlassScaledMMLinearKernel],
-    PlatformEnum.ROCM: [TritonScaledMMLinearKernel],
+    PlatformEnum.ROCM: [AiterScaledMMLinearKernel, TritonScaledMMLinearKernel],
     PlatformEnum.TPU: [XLAScaledMMLinearKernel],
 }
 

vllm/model_executor/layers/quantization/kernels/scaled_mm/aiter.py

@@ -0,0 +1,119 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Optional, Tuple
+
+import torch
+
+import vllm.envs as envs
+from vllm import _custom_ops as ops
+from vllm.platforms import current_platform
+
+from .cutlass import CutlassScaledMMLinearKernel
+from .ScaledMMLinearKernel import ScaledMMLinearLayerConfig
+
+
+class AiterScaledMMLinearKernel(CutlassScaledMMLinearKernel):
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        return 90
+
+    @classmethod
+    def can_implement(
+            cls, c: ScaledMMLinearLayerConfig) -> Tuple[bool, Optional[str]]:
+        if not current_platform.is_rocm():
+            return (
+                False,
+                "AiterScaledMMLinearKernel requires `aiter` which is not " +
+                "currently supported on non-ROCm platform.")
+
+        try:
+            import aiter  # noqa: F401 # deliberately attempt to import aiter
+        except Exception:
+            return (
+                False,
+                "AiterScaledMMLinearKernel requires `aiter` which is not " +
+                "installed on ROCm.")
+        # Check if rocm_aiter_gemm_w8a8_scaled_mm is enabled
+        if not (
+            envs.VLLM_ROCM_USE_AITER_LINEAR \
+            and envs.VLLM_ROCM_USE_AITER
+        ):
+            return (False, "AiterScaledMMLinearKernel is disabled. " +
+                    "Enable by setting `VLLM_ROCM_USE_AITER=1` " +
+                    "and `VLLM_ROCM_USE_AITER_LINEAR=1`. " +
+                    "`VLLM_ROCM_USE_AITER_LINEAR` default is True.")
+
+        if not c.input_symmetric:
+            return (False,
+                    "AiterScaledMMLinearKernel only supports symmetric " +
+                    "quantization.")
+        return True, None
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        super().process_weights_after_loading(layer)
+
+    def apply_weights(self,
+                      layer: torch.nn.Module,
+                      x: torch.Tensor,
+                      bias: Optional[torch.Tensor] = None) -> torch.Tensor:
+        """
+        `AiterScaledMMLinearKernel` implements a fused version of
+            `output = torch.mm((scale_a * a), (scale_b * b)).to(out_dtype)`
+        where scale_a * a and scale_b * b are implemented using numpy-style
+        broadcasting.
+        Currently only support per-tensor-per-tensor GEMM
+        and per-token-per-channel GEMM through AITER
+        w8a8 scaled gemm. `AiterScaledMMLinearKernel` also does not support
+        ATIER block scaled GEMM and mix-precision GEMM.
+        """
+        w_q, w_s, i_s, i_zp, azp_adj = self._get_weight_params(layer)
+
+        # ops.scaled_int8_quant supports both dynamic and static quant:
+        # * dynamic, i_s is None and x_s computed from x.
+        # * static, i_s is scalar and x_s is i_s.
+        symmetric = azp_adj is None
+        assert symmetric, ("AiterScaledMMLinearKernel only supports"
+                           " symmetric quantization.")
+        x_q, x_s, x_zp = ops.scaled_int8_quant(x,
+                                               i_s,
+                                               i_zp,
+                                               symmetric=symmetric)
+
+        assert x_zp is None, ("AiterScaledMMLinearKernel only supports"
+                              " symmetric quantization.")
+        out_dtype = x.dtype
+
+        assert (w_q.shape[0] % 16 == 0 and w_q.shape[1] % 16 == 0)
+        assert (out_dtype is torch.bfloat16 or out_dtype is torch.float16)
+        assert bias is None or bias.shape[0] == w_q.shape[
+            1] and bias.dtype == out_dtype
+
+        m = x_q.shape[0]  # a
+        n = w_q.shape[1]  # b
+
+        per_tensor_scale_a = (x_s.numel() == 1)
+        per_tensor_scale_b = (w_s.numel() == 1)
+        per_token_scale_a = (x_s.numel() == m)
+        per_channel_scale_b = (w_s.numel() == n)
+
+        # @TODO:
+        # Maybe broadcast the per-tensor-scale into per-channel-scale
+        # if one of the scale is a per-channel-scale.
+        # For now, it only supports:
+        # - per-tensor-per-tensor a8w8 scaled GEMM, and
+        # - per-token-per-channel a8w8 scaled GEMM
+        assert ((per_tensor_scale_a and per_tensor_scale_b)
+                or (per_token_scale_a and per_channel_scale_b)), (
+                    "Currently only support per-tensor-per-tensor GEMM " +
+                    " and per-token-per-channel GEMM through AITER"
+                    " w8a8 scaled gemm. `AiterScaledMMLinearKernel` " +
+                    "does not support AITER block scaled GEMM.")
+
+        from aiter import gemm_a8w8_CK
+
+        # gemm_a8w8_CK(a, b, scale_a, scale_b, bias) expects
+        # a to be [M, K]
+        # b to be [N, K]
+        # CutlassScaledMMLinearKernel prepare weight `w_q` in [K, N] format
+        return gemm_a8w8_CK(x_q, w_q.t(), x_s, w_s, bias).to(out_dtype)