# PR #53161 完整报告

- 仓库：`vllm-project/vllm`
- 标题：[ROCm][Perf][DeepSeek V4] Fuse native FP8 shared expert with MXFP4 routed experts
- 合并时间：2026-09-07 10:42
- 原文链接：http://prhub.com.cn/vllm-project/vllm/pull/53161

---

# 执行摘要

- 一句话：
- 推荐动作：

# 功能与动机



# 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

关键文件：
- `vllm/models/deepseek_v4/amd/model.py`（模块 vllm/models；类别 source；类型 data-contract；符号 _use_heterogeneous_fhmoe, _validate_heterogeneous_routes, _heterogeneous_shared_expert_enabled, _prepare_native_fp8_shared_expert）: 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
- `vllm/_aiter_ops.py`（模块 vllm；类别 source；类型 dependency-wiring；符号 _validate_rocm_aiter_fused_moe_shared_expert_args, _probe_dsv4_i384_fhmoe_capability, fused_moe_supports_heterogeneous_shared_expert, shuffle_scale）: 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
- `tests/model_executor/layers/test_fused_shared_expert.py`（模块 fused/shared/expert；类别 test；类型 test-coverage；符号 test_deepseek_v4_heterogeneous_fhmoe_keeps_native_intermediate_width, test_deepseek_v4_heterogeneous_fhmoe_token_policy, supports, _supported_fhmoe_signature）: 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
- `vllm/model_executor/layers/fused_moe/experts/rocm_aiter_moe.py`（模块 vllm/model_executor；类别 source；类型 data-contract）: 源码主路径；包含 控制流调整、配置键调整；+15/-0

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


# 评论区精华



- 暂无高价值评论线程

# 风险与影响

- 风险：
- 影响：
 - 风险标记：暂无

# 关联脉络

- 暂无明显关联 PR