[Mini-VLLM] Remove the ROCm device path, debug routes and arch-gated tests
- Delete csrc/rocm, hipify, the _rocm_C extension and the setup.py rocm target path; drop the rocm requirement files - Delete the aiter, rdna3, skinny-gemm and ROCm fused-MoE kernel modules and retarget the linear kernel registries away from them - Reduce platforms/rocm.py to the GCN architecture predicates that live code still imports, and delete the RocmPlatform class and plugin - Drop the VLLM_SERVER_DEV_MODE debug surface: reset, collective_rpc, server_info and sleep routes plus the log_error_stack flag - Delete lm-format-enforcer, lark and otel-semantic-conventions-ai pins; the live server process maps none of them - Delete tests that only build architectures this fork no longer registers, and test directories left without any test file - Restore the tests/v1/kv_connector mock helpers that scheduler tests import at module level Co-authored-by: Hinata <hinata@idihore.id> Signed-off-by: MOVZX <movzx@yahoo.com>
This commit is contained in:
@@ -28,9 +28,6 @@ from vllm.model_executor.kernels.linear.scaled_mm.pytorch import (
|
||||
PerTensorTorchFP8ScaledMMLinearKernel,
|
||||
RowWiseTorchFP8ScaledMMLinearKernel,
|
||||
)
|
||||
from vllm.model_executor.kernels.linear.scaled_mm.rocm import (
|
||||
ROCmFP8ScaledMMLinearKernel,
|
||||
)
|
||||
from vllm.model_executor.kernels.linear.scaled_mm.ScaledMMLinearKernel import (
|
||||
FP8ScaledMMLinearKernel,
|
||||
FP8ScaledMMLinearLayerConfig,
|
||||
@@ -62,7 +59,6 @@ __all__ = [
|
||||
"ChannelWiseTorchFP8ScaledMMLinearKernel",
|
||||
"PerTensorTorchFP8ScaledMMLinearKernel",
|
||||
"RowWiseTorchFP8ScaledMMLinearKernel",
|
||||
"ROCmFP8ScaledMMLinearKernel",
|
||||
"TritonFP8ScaledMMLinearKernel",
|
||||
"TritonInt8ScaledMMLinearKernel",
|
||||
"ZentorchInt8ScaledMMLinearKernel",
|
||||
|
||||
@@ -1,123 +0,0 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
|
||||
import torch
|
||||
|
||||
import vllm.envs as envs
|
||||
from vllm import _custom_ops as ops
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.utils.platform_utils import num_compute_units
|
||||
from vllm.utils.torch_utils import direct_register_custom_op
|
||||
|
||||
from .ScaledMMLinearKernel import (
|
||||
FP8ScaledMMLinearKernel,
|
||||
FP8ScaledMMLinearLayerConfig,
|
||||
)
|
||||
|
||||
|
||||
def rocm_per_tensor_float_w8a8_scaled_mm_impl(
|
||||
A: torch.Tensor,
|
||||
B: torch.Tensor,
|
||||
out_dtype: torch.dtype,
|
||||
As: torch.Tensor,
|
||||
Bs: torch.Tensor,
|
||||
bias: torch.Tensor,
|
||||
) -> torch.Tensor:
|
||||
if (
|
||||
A.shape[0] <= 4
|
||||
and B.shape[0] % 16 == 0 # M TODO: needed?
|
||||
and B.shape[1] % 16 == 0 # K
|
||||
and ((bias is None) or (bias.dtype == out_dtype))
|
||||
):
|
||||
output = ops.wvSplitKQ(
|
||||
B.t(),
|
||||
A,
|
||||
out_dtype,
|
||||
As,
|
||||
Bs,
|
||||
num_compute_units(),
|
||||
bias,
|
||||
)
|
||||
# Fallback
|
||||
else:
|
||||
output = torch._scaled_mm(
|
||||
A,
|
||||
B,
|
||||
out_dtype=out_dtype,
|
||||
scale_a=As,
|
||||
scale_b=Bs,
|
||||
bias=bias,
|
||||
)
|
||||
return output
|
||||
|
||||
|
||||
def rocm_per_tensor_float_w8a8_scaled_mm_fake(
|
||||
A: torch.Tensor,
|
||||
B: torch.Tensor,
|
||||
out_dtype: torch.dtype,
|
||||
As: torch.Tensor,
|
||||
Bs: torch.Tensor,
|
||||
bias: torch.Tensor,
|
||||
) -> torch.Tensor:
|
||||
return A.new_empty((*A.shape[:-1], B.shape[1]), dtype=out_dtype)
|
||||
|
||||
|
||||
if current_platform.is_rocm():
|
||||
direct_register_custom_op(
|
||||
op_name="rocm_per_tensor_float_w8a8_scaled_mm_impl",
|
||||
op_func=rocm_per_tensor_float_w8a8_scaled_mm_impl,
|
||||
fake_impl=rocm_per_tensor_float_w8a8_scaled_mm_fake,
|
||||
)
|
||||
|
||||
|
||||
class ROCmFP8ScaledMMLinearKernel(FP8ScaledMMLinearKernel):
|
||||
@classmethod
|
||||
def is_supported(
|
||||
cls, compute_capability: int | None = None
|
||||
) -> tuple[bool, str | None]:
|
||||
if not current_platform.is_rocm():
|
||||
return False, "requires ROCm."
|
||||
|
||||
from vllm.platforms.rocm import get_cdna_version, on_gfx12x, on_gfx1250
|
||||
|
||||
# wvSplitKQ (skinny GEMM) is excluded from the gfx1250 build.
|
||||
if on_gfx1250():
|
||||
return False, "wvSplitKQ (skinny GEMM) is not built on gfx1250"
|
||||
|
||||
# Restore RDNA4 (gfx12x) dropped by the get_cdna_version()>2 refactor.
|
||||
if get_cdna_version() <= 2 and not on_gfx12x():
|
||||
return False, "requires CDNA3+ (gfx942/gfx950) or RDNA4 (gfx12x)"
|
||||
|
||||
if not envs.VLLM_ROCM_USE_SKINNY_GEMM:
|
||||
return False, "requires VLLM_ROCM_USE_SKINNY_GEMM to be enabled."
|
||||
|
||||
return True, None
|
||||
|
||||
@classmethod
|
||||
def can_implement(cls, c: FP8ScaledMMLinearLayerConfig) -> tuple[bool, str | None]:
|
||||
per_tensor_activation_scales = (
|
||||
c.activation_quant_key.scale.group_shape.is_per_tensor()
|
||||
)
|
||||
per_tensor_weight_scales = c.weight_quant_key.scale.group_shape.is_per_tensor()
|
||||
|
||||
if not (per_tensor_activation_scales and per_tensor_weight_scales):
|
||||
return False, "requires per tensor activation and weight scales."
|
||||
|
||||
return True, None
|
||||
|
||||
def apply_scaled_mm(
|
||||
self,
|
||||
*,
|
||||
A: torch.Tensor,
|
||||
B: torch.Tensor,
|
||||
out_dtype: torch.dtype,
|
||||
As: torch.Tensor,
|
||||
Bs: torch.Tensor,
|
||||
bias: torch.Tensor | None,
|
||||
output_shape: list,
|
||||
) -> torch.Tensor:
|
||||
output = torch.ops.vllm.rocm_per_tensor_float_w8a8_scaled_mm_impl(
|
||||
A, B, out_dtype, As, Bs, bias
|
||||
)
|
||||
return torch.narrow(output, 0, 0, A.shape[0]).view(*output_shape)
|
||||
Reference in New Issue
Block a user