Claude
|
6415249693
|
data: port complete MoE + xllm layer call chains from upstream repos
MoE call chain from ds_vllm (vllm-project/vllm latest):
ex_engine/moe/ — 20 files, 8736 lines
- modular_kernel.py (1630 lines) — base classes for modular MoE
- experts/fused_batched_moe.py (972 lines) — NaiveBatchedExperts
- prepare_finalize/batched.py (171 lines) — token grouping by expert
- topk_weight_and_reduce.py (176 lines) — scatter-add finalize
- fused_moe.py (1740 lines) — main fused_moe dispatch
- config.py (1407 lines) — FusedMoEQuantConfig
- activation.py, utils.py, layer.py, etc.
xllm layer code (jd-opensource/xllm):
ex_engine/xllm_layers/ — 39 files, 5859 lines
- ilu/fused_moe.cpp (797 lines) — production ixformer 7-step MoE pipeline
- ilu/attention.cpp (189 lines) — paged_attention + flash_attn bridge
- npu_torch/qwen3_gated_delta_net_base.cpp (576 lines) — GDN reference
- common/rms_norm.cpp, rotary_embedding.cpp, activation.cpp, dense_mlp.cpp
xllm ILU kernels — synced 10 files to upstream (diffs from prior edits)
These are reference implementations, NOT hand-written.
Source repos: vllm-project/vllm, jd-opensource/xllm
|
2026-08-15 14:26:24 +00:00 |
|
dylan
|
e18ece8f3a
|
feat: port NaiveBatchedExperts from ds_vllm — view transpose + cublas transB
Source: upstream_ref/ds_vllm/vllm/model_executor/layers/fused_moe/experts/fused_batched_moe.py
upstream_ref/ds_vllm/vllm/model_executor/layers/fused_moe/activation.py
New files (ported from ds_vllm, adapted for BI-V100):
ex_engine/moe/__init__.py
ex_engine/moe/activation.py
- MoEActivation enum + apply_moe_activation
- torch.ops._C.silu_and_mul replaced with F.silu(gate)*up fallback
ex_engine/moe/naive_batched_experts.py
- naive_batched_moe_forward()
- Decode: per-expert loop, w13[eid].transpose(0,1) is VIEW (zero copy)
- @ operator → cublas passes transB=CUBLAS_OP_T internally
- Prefill: group tokens by expert, batch @ per expert
Modified:
qwen3_6_scripts/qwen3_5.py
- Import naive_batched_moe_forward
- Tier 0.5: after ix_fused_moe, before corex point-optimized loop
- Uses existing topk routing (xllm/corex/pytorch)
Key difference from previous approach:
- NO physical transpose (was 22ms overhead)
- NO weight gather into contiguous buffer
- View transpose is O(0), cublas handles transB
|
2026-08-15 13:05:45 +00:00 |
|