[Refactor] Remove Duplicate per_block_cast_to_fp8, Remove Dependencies of DeepGEMM (#21787)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Wentao Ye
2025-08-01 01:13:27 +00:00
committed by GitHub
parent 0bd409cf01
commit 3700642013
8 changed files with 55 additions and 132 deletions
@@ -1,10 +1,10 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import math
import torch
import vllm._custom_ops as ops
from vllm.utils.deep_gemm import per_block_cast_to_fp8
def per_token_cast_to_fp8(
@@ -20,29 +20,6 @@ def per_token_cast_to_fp8(
return fp8_data.view(m, n + pad_size)[:, :n], (x_amax / 448.0).view(m, -1)
def per_block_cast_to_fp8(
x: torch.Tensor, block_size_k: int,
block_size_n: int) -> tuple[torch.Tensor, torch.Tensor]:
assert x.dim() == 2
m, n = x.shape
x_padded = torch.zeros(
(
int(math.ceil(m / block_size_k)) * block_size_k,
int(math.ceil(n / block_size_n)) * block_size_n,
),
dtype=x.dtype,
device=x.device,
)
x_padded[:m, :n] = x
x_view = x_padded.view(-1, block_size_k,
x_padded.size(1) // block_size_k, block_size_n)
x_amax = x_view.abs().float().amax(dim=(1, 3), keepdim=True).clamp(1e-4)
x_scaled = (x_view * (448.0 / x_amax)).to(torch.float8_e4m3fn)
x_scaled_sub = x_scaled.view_as(x_padded)[:m, :n].contiguous()
scales = (x_amax / 448.0).view(x_view.size(0), x_view.size(2))
return x_scaled_sub, scales
def make_non_quant_weights(
e: int,
n: int,
@@ -99,11 +76,9 @@ def make_block_quant_fp8_weights(
for i in range(e):
w1[i], w1_s[i] = per_block_cast_to_fp8(w1_bf16[i],
block_size_k=block_k,
block_size_n=block_n)
block_size=[block_k, block_n])
w2[i], w2_s[i] = per_block_cast_to_fp8(w2_bf16[i],
block_size_k=block_k,
block_size_n=block_n)
block_size=[block_k, block_n])
return w1, w2, w1_s, w2_s
+2 -19
View File
@@ -12,10 +12,8 @@ import torch
from tests.kernels.utils import baseline_scaled_mm
from vllm import _custom_ops as ops
from vllm.platforms import current_platform
def cdiv(a, b):
return (a + b - 1) // b
from vllm.utils import cdiv
from vllm.utils.deep_gemm import per_block_cast_to_fp8
def per_token_cast_to_fp8(
@@ -32,21 +30,6 @@ def per_token_cast_to_fp8(
return fp8_data.view(m, n + pad_size)[:, :n], (x_amax / 448.0).view(m, -1)
def per_block_cast_to_fp8(
x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
assert x.dim() == 2
m, n = x.shape
x_padded = torch.zeros((cdiv(m, 128) * 128, cdiv(n, 128) * 128),
device=x.device,
dtype=x.dtype)
x_padded[:m, :n] = x
x_view = x_padded.view(-1, 128, x_padded.size(1) // 128, 128)
x_amax = x_view.abs().float().amax(dim=(1, 3), keepdim=True).clamp(1e-4)
x_scaled = (x_view * (448.0 / x_amax)).to(dtype=torch.float8_e4m3fn)
return x_scaled.view_as(x_padded)[:m, :n].contiguous(), (
x_amax / 448.0).view(x_view.size(0), x_view.size(2))
@pytest.mark.parametrize("num_groups, expected_m_per_group, k, n", [
(4, 8192, 7168, 4096),
(4, 8192, 2048, 7168),
+6 -2
View File
@@ -69,8 +69,12 @@ def make_block_quant_fp8_weights(
dtype=torch.float32)
for i in range(e):
w1[i], w1_s[i] = per_block_cast_to_fp8(w1_bf16[i])
w2[i], w2_s[i] = per_block_cast_to_fp8(w2_bf16[i])
w1[i], w1_s[i] = per_block_cast_to_fp8(w1_bf16[i],
block_size=block_size,
use_ue8m0=True)
w2[i], w2_s[i] = per_block_cast_to_fp8(w2_bf16[i],
block_size=block_size,
use_ue8m0=True)
return w1, w2, w1_s, w2_s
+2 -2
View File
@@ -5,8 +5,7 @@ from typing import Optional
import torch
import vllm._custom_ops as ops
from tests.kernels.quant_utils import (per_block_cast_to_fp8,
per_block_cast_to_int8)
from tests.kernels.quant_utils import per_block_cast_to_int8
from vllm.model_executor.layers.fused_moe import fused_experts
from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
BatchedPrepareAndFinalize, BatchedTritonExperts, NaiveBatchedExperts)
@@ -15,6 +14,7 @@ from vllm.model_executor.layers.fused_moe.modular_kernel import (
from vllm.model_executor.layers.fused_moe.utils import (
moe_kernel_quantize_input)
from vllm.utils import round_up
from vllm.utils.deep_gemm import per_block_cast_to_fp8
def triton_moe(
-19
View File
@@ -222,25 +222,6 @@ def native_per_token_group_quant_int8(x,
DEFAULT_BLOCK_SHAPE = [128, 128]
def per_block_cast_to_fp8(
x: torch.Tensor,
block_shape: list[int] = DEFAULT_BLOCK_SHAPE,
) -> tuple[torch.Tensor, torch.Tensor]:
block_m, block_n = block_shape
assert x.dim() == 2
m, n = x.shape
x_padded = torch.zeros((round_up(m, block_m), round_up(n, block_n)),
dtype=x.dtype,
device=x.device)
x_padded[:m, :n] = x
x_view = x_padded.view(-1, block_m, x_padded.size(1) // block_n, block_n)
x_amax = x_view.abs().float().amax(dim=(1, 3), keepdim=True).clamp(1e-4)
x_scaled = (x_view * (448.0 / x_amax)).to(torch.float8_e4m3fn)
x_scaled_sub = x_scaled.view_as(x_padded)[:m, :n].contiguous()
scales = (x_amax / 448.0).view(x_view.size(0), x_view.size(2))
return x_scaled_sub, scales
def per_block_cast_to_int8(
x: torch.Tensor,
block_shape: list[int] = DEFAULT_BLOCK_SHAPE,
+1 -1
View File
@@ -117,7 +117,7 @@ def test_w8a8_block_fp8_deep_gemm_matmul(M, N, K, block_size, out_dtype, seed):
B_fp32 = (torch.rand(N, K, dtype=torch.float32) - 0.5) * 2 * fp8_max
A_fp8, As_fp8 = per_token_group_quant_fp8(A_fp32, block_size[1])
B_fp8, Bs_fp8 = per_block_cast_to_fp8(B_fp32)
B_fp8, Bs_fp8 = per_block_cast_to_fp8(B_fp32, block_size=block_size)
As = As_fp8.to(torch.float32)
Bs = Bs_fp8.to(torch.float32)