Prhub

#53161 [ROCm][Perf][DeepSeek V4] Fuse native FP8 shared expert with MXFP4 routed experts

原始 PR 作者 Fangzhou-Ai 合并时间 2026-09-07 10:42 文件变更 4 提交数 7 评论 15 代码增减 +839 / -6

执行摘要

实现拆解

  1. 核心源码路径
    - vllm/models/deepseek_v4/amd/model.py(data-contract):源码主路径;涉及符号 _use_heterogeneous_fhmoe, _validate_heterogeneous_routes, _heterogeneous_shared_expert_enabled, _prepare_native_fp8_shared_expert;包含 导入关系调整、控制流调整、配置键调整;+377/-5;关键符号:_use_heterogeneous_fhmoe, _validate_heterogeneous_routes, _heterogeneous_shared_expert_enabled, _prepare_native_fp8_shared_expert
    - vllm/_aiter_ops.py(dependency-wiring):源码主路径;涉及符号 _validate_rocm_aiter_fused_moe_shared_expert_args, _probe_dsv4_i384_fhmoe_capability, fused_moe_supports_heterogeneous_shared_expert, shuffle_scale;包含 导入关系调整、控制流调整、配置键调整;+107/-0;关键符号:_validate_rocm_aiter_fused_moe_shared_expert_args, _probe_dsv4_i384_fhmoe_capability, fused_moe_supports_heterogeneous_shared_expert, shuffle_scale
    - vllm/model_executor/layers/fused_moe/experts/rocm_aiter_moe.py(data-contract):源码主路径;包含 控制流调整、配置键调整;+15/-0

  2. 测试与验证
    - tests/model_executor/layers/test_fused_shared_expert.py(test-coverage):测试配套;涉及符号 test_deepseek_v4_heterogeneous_fhmoe_keeps_native_intermediate_width, test_deepseek_v4_heterogeneous_fhmoe_token_policy, supports, _supported_fhmoe_signature;包含 测试覆盖调整、导入关系调整、控制流调整;+340/-1;关键符号:test_deepseek_v4_heterogeneous_fhmoe_keeps_native_intermediate_width, test_deepseek_v4_heterogeneous_fhmoe_token_policy, supports, _supported_fhmoe_signature

文件 模块 状态 重要度
vllm/models/deepseek_v4/amd/model.py vllm/models modified 9.21
vllm/_aiter_ops.py vllm modified 8.55
tests/model_executor/layers/test_fused_shared_expert.py fused/shared/expert modified 8.1
vllm/model_executor/layers/fused_moe/experts/rocm_aiter_moe.py vllm/model_executor modified 6.37

关键符号

_use_heterogeneous_fhmoe _validate_heterogeneous_routes _heterogeneous_shared_expert_enabled _prepare_native_fp8_shared_expert scale_bytes DeepseekV4HeterogeneousMxfp4MoEMethod process_weights_after_loading DeepseekV4HeterogeneousSharedRoutedExperts _validate_rocm_aiter_fused_moe_shared_expert_args _probe_dsv4_i384_fhmoe_capability

分析完成后,这里会展示 LLM 生成的相对完整源码片段和详细注释。

评论区精华

没有提炼出高价值讨论线程

当前评论区没有形成足够清晰的争议点或结论,后续有更多讨论时会体现在这里。

风险与影响

关联 Issue

#4891 [FlyDSL] [JIT] [AOT] [FHMoE] Support native-I384 DeepSeek V4 through M=2048
#48728 [ROCm][Perf][DeepSeek V4] Enable heterogeneous MXFP4/FP8 fused shared expert

完整报告

参与讨论