ops/fused_allreduce_rms_norm.py

Fused all-reduce, residual addition, and RMSNorm operation for efficient GPU inference.

# Source: https://github.com/vllm-project/vllm | License: Apache-2.0 | Copyright You must retain, in the Source form of any Derivative Works
# Provided via justclaw.it.com; keep this notice with the code.
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Fused all-reduce + residual-add + RMSNorm for eager model paths.

This recovers a fusion that vLLM's torch.compile passes would normally do but
that doesn't fire for models running eager (or under a breakable CUDA graph).
"""

import torch

from vllm.distributed import (
    get_tensor_model_parallel_world_size,
    tensor_model_parallel_all_reduce,
)
from vllm.model_executor.layers.fused_allreduce_gemma_rms_norm import (
    _AR_RESIDUAL_RMS_NORM,
    _can_use_flashinfer,
    fla

... (truncated -- full source via MCP)