Add NVFP4 all-gather GEMM fusion for AsyncTP (#41882)

Signed-off-by: roG0d <baonudesifeizhai@gmail.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
baonudesifeizhai
2026-05-10 01:13:22 +00:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 006af4b956
commit bc5fdc1e6a
7 changed files with 605 additions and 6 deletions
+2 -1
View File
@@ -17,7 +17,6 @@ import pytest
import torch
from torch import nn
from vllm.compilation.counter import compilation_counter
from vllm.compilation.decorators import support_torch_compile
from vllm.config import (
CompilationConfig,
@@ -340,6 +339,8 @@ def run_model(llama_config, compile_config: CompilationConfig) -> torch.Tensor:
def test_toy_llama(
backend: str, use_inductor_graph_partition: bool, monkeypatch, tmp_path
):
from vllm.compilation.counter import compilation_counter
# We disable the vLLM compile cache into a new tmp dir for 1 reason:
# 1. To make sure we can properly track the number of Inductor compilations.
monkeypatch.setenv("VLLM_DISABLE_COMPILE_CACHE", "1")