[Performance] Move apply_w8a8_block_fp8_linear to an op class (#24666)

Signed-off-by: ElizaWszola <ewszola@redhat.com>
Signed-off-by: ElizaWszola <elizaw.9289@gmail.com>
Signed-off-by: Luka Govedič <lgovedic@redhat.com>
Signed-off-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
Co-authored-by: Luka Govedič <lgovedic@redhat.com>
This commit is contained in:
ElizaWszola
2025-09-23 12:03:10 -07:00
committed by GitHub
co-authored by Luka Govedič Michael Goin Luka Govedič
parent 8c1c81a3de
commit 63400259d0
14 changed files with 341 additions and 201 deletions
+3 -2
View File
@@ -12,7 +12,7 @@ from tests.kernels.quant_utils import (native_per_token_group_quant_fp8,
from vllm.config import VllmConfig
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
cutlass_scaled_mm, get_col_major_tma_aligned_tensor,
per_token_group_quant_fp8, w8a8_block_fp8_matmul)
per_token_group_quant_fp8, w8a8_triton_block_scaled_mm)
from vllm.platforms import current_platform
from vllm.utils import has_deep_gemm
from vllm.utils.deep_gemm import fp8_gemm_nt, per_block_cast_to_fp8
@@ -90,7 +90,8 @@ def test_w8a8_block_fp8_matmul(M, N, K, block_size, out_dtype, seed):
ref_out = native_w8a8_block_matmul(A_fp8, B_fp8, As, Bs, block_size,
out_dtype)
out = w8a8_block_fp8_matmul(A_fp8, B_fp8, As, Bs, block_size, out_dtype)
out = w8a8_triton_block_scaled_mm(A_fp8, B_fp8, As, Bs, block_size,
out_dtype)
rel_diff = (torch.mean(
torch.abs(out.to(torch.float32) - ref_out.to(torch.float32))) /
@@ -20,9 +20,11 @@ from vllm.platforms import current_platform
(8, 513, 64), # Non-divisible (native only)
])
@pytest.mark.parametrize("seed", [42])
@pytest.mark.parametrize("use_ue8m0", [True, False])
@torch.inference_mode()
def test_quantfp8_group_functionality(batch_size: int, hidden_dim: int,
group_size: int, seed: int) -> None:
group_size: int, seed: int,
use_ue8m0: bool) -> None:
"""Test QuantFP8 group quantization with various configurations.
Tests both CUDA and native implementations, column-major scales,
@@ -38,7 +40,8 @@ def test_quantfp8_group_functionality(batch_size: int, hidden_dim: int,
group_shape = GroupShape(1, group_size)
quant_op = QuantFP8(static=False,
group_shape=group_shape,
column_major_scales=False)
column_major_scales=False,
use_ue8m0=use_ue8m0)
# 1. Test native implementation (always available)
x_quant_native, scales_native = quant_op.forward_native(x.clone())
@@ -48,9 +51,15 @@ def test_quantfp8_group_functionality(batch_size: int, hidden_dim: int,
# 2. Test column-major scales configuration
quant_op_col = QuantFP8(static=False,
group_shape=group_shape,
column_major_scales=True)
column_major_scales=True,
use_ue8m0=use_ue8m0)
_, scales_col = quant_op_col.forward_native(x.clone())
assert scales_col.shape == (expected_num_groups, batch_size)
assert scales_col.shape == (batch_size, expected_num_groups)
assert scales_col.stride(0) == 1
assert scales_col.stride(1) == batch_size
# Test column-major scales consistency
assert torch.allclose(scales_col, scales_native, rtol=1e-9, atol=1e-8)
# 3. Test CUDA implementation (only for divisible dimensions)
if is_divisible:
@@ -68,8 +77,9 @@ def test_quantfp8_group_functionality(batch_size: int, hidden_dim: int,
@pytest.mark.parametrize("seed", [42])
@pytest.mark.parametrize("use_ue8m0", [True, False])
@torch.inference_mode()
def test_quantfp8_group_multidimensional(seed: int) -> None:
def test_quantfp8_group_multidimensional(seed: int, use_ue8m0: bool) -> None:
current_platform.seed_everything(seed)
group_size = 64
@@ -82,7 +92,8 @@ def test_quantfp8_group_multidimensional(seed: int) -> None:
group_shape = GroupShape(1, group_size)
quant_op = QuantFP8(static=False,
group_shape=group_shape,
column_major_scales=False)
column_major_scales=False,
use_ue8m0=use_ue8m0)
x_quant, scales = quant_op.forward_native(x_3d.clone())
assert x_quant.shape == x_3d.shape
@@ -91,7 +102,8 @@ def test_quantfp8_group_multidimensional(seed: int) -> None:
# Test column_major_scales with multi-dim
quant_op_col = QuantFP8(static=False,
group_shape=group_shape,
column_major_scales=True)
column_major_scales=True,
use_ue8m0=use_ue8m0)
_, scales_col = quant_op_col.forward_native(x_3d.clone())
assert scales_col.shape == (batch1, hidden_dim // group_size, batch2)
@@ -17,8 +17,6 @@ from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
from vllm.model_executor.layers.layernorm import (RMSNorm,
dispatch_rocm_rmsnorm_func,
fused_add_rms_norm, rms_norm)
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
cutlass_scaled_mm, dispatch_w8a8_blockscale_func, w8a8_block_fp8_matmul)
from vllm.platforms import current_platform
RMS_NORM_SUPPORTED_DTYPES = [torch.float16, torch.bfloat16]
@@ -111,34 +109,6 @@ def test_enabled_ops_invalid(env: str):
RMSNorm(1024).enabled()
@pytest.mark.skipif(
not current_platform.is_rocm() or not current_platform.is_fp8_fnuz(),
reason="AITER is a feature exclusive for ROCm and FP8_FNUZ")
@pytest.mark.parametrize("use_cutlass", [True, False])
@pytest.mark.parametrize("use_rocm_aiter", ["0", "1"])
@pytest.mark.parametrize("use_rocm_aiter_gemm_w8a8_blockscale", ["0", "1"])
def test_w8a8_blockscale_dispatch(use_cutlass: bool, use_rocm_aiter: str,
use_rocm_aiter_gemm_w8a8_blockscale: str,
monkeypatch):
monkeypatch.setenv("VLLM_ROCM_USE_AITER", use_rocm_aiter)
monkeypatch.setenv("VLLM_ROCM_USE_AITER_LINEAR",
use_rocm_aiter_gemm_w8a8_blockscale)
use_aiter_and_is_supported = (bool(int(use_rocm_aiter)) and bool(
int(use_rocm_aiter_gemm_w8a8_blockscale)))
block_scale_func = dispatch_w8a8_blockscale_func(
use_cutlass, use_aiter_and_is_supported=use_aiter_and_is_supported)
if use_cutlass:
assert block_scale_func == cutlass_scaled_mm
elif current_platform.is_rocm() and int(use_rocm_aiter) and int(
use_rocm_aiter_gemm_w8a8_blockscale):
assert block_scale_func == (
torch.ops.vllm.rocm_aiter_gemm_w8a8_blockscale)
else:
assert block_scale_func == w8a8_block_fp8_matmul
@pytest.mark.parametrize("use_rocm_aiter", ["0", "1"])
def test_topk_dispatch(use_rocm_aiter: str, monkeypatch):
monkeypatch.setenv("VLLM_ROCM_USE_AITER", use_rocm_aiter)
@@ -18,6 +18,9 @@ from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tenso
CompressedTensorsW4A16Fp4, CompressedTensorsW4A16Sparse24,
CompressedTensorsW8A8Fp8, CompressedTensorsW8A8Int8,
CompressedTensorsW8A16Fp8, CompressedTensorsWNA16)
from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
W8A8BlockFp8LinearOp)
from vllm.model_executor.layers.quantization.utils.quant_utils import (
cutlass_fp4_supported)
from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
@@ -742,3 +745,35 @@ def test_compressed_tensors_transforms_perplexity(vllm_runner, model, prompt,
perplexity = llm.generate_prompt_perplexity([prompt])[0]
print(perplexity)
assert perplexity <= exp_perplexity
def test_compressed_tensors_fp8_block_enabled(vllm_runner):
model_path = "RedHatAI/Qwen3-0.6B-FP8-BLOCK"
with vllm_runner(model_path) as llm:
fp8_dtype = current_platform.fp8_dtype()
def check_model(model):
layer = model.model.layers[0]
qkv_proj = layer.self_attn.qkv_proj
assert isinstance(qkv_proj.quant_method,
CompressedTensorsLinearMethod)
assert isinstance(qkv_proj.scheme, CompressedTensorsW8A8Fp8)
assert isinstance(qkv_proj.scheme.w8a8_block_fp8_linear,
W8A8BlockFp8LinearOp)
assert qkv_proj.weight.dtype is fp8_dtype
assert qkv_proj.weight_scale.dtype is torch.float32
assert len(qkv_proj.weight.shape) == 2
assert len(qkv_proj.weight_scale.shape) == 2
input_quant_op = \
qkv_proj.scheme.w8a8_block_fp8_linear.input_quant_op
assert isinstance(input_quant_op, QuantFP8)
assert input_quant_op._forward_method == input_quant_op.forward_cuda
llm.apply_model(check_model)
output = llm.generate_greedy("Hello my name is", max_tokens=20)
assert output