project6-dev f28223c9da perf: native ixformer decode (v1 ≤32K, v2 >32K) + flash_attn_varlen prefill ...

Replaces all Python PyTorch fallback attention with native ixformer kernels:

Decode path:
- ≤32K: paged_attention_v1 (5D KV layout, x=8) — verified on real BI-V100
- >32K: paged_attention_v2 (5D→4D permute) — verified 65K+ on real BI-V100
- Removes _forward_decode_pytorch Python fallback entirely

Prefill path (profiling):
- _run_sdpa_fallback now uses ixformer.flash_attn_varlen_func
- head_dim=256 verified correct (diff<0.004) and 1.7x faster than PyTorch
- Falls back to Q-tiling pure-math if ixformer unavailable

Also includes: MoE kernel integration, GDN C++ kernels, diagnostic scripts,
xllm upstream layer/kernel references, .dockerignore cleanup.

All changes verified on real BI-V100 hardware (single card).

2026-08-13 07:04:21 +00:00

..

ilu

fix(build): 回退qwen3_6_scripts+ex_engine到26e6cb40(能得分版本)

2026-08-12 01:33:24 +00:00

moe

fix(build): 回退qwen3_6_scripts+ex_engine到26e6cb40(能得分版本)

2026-08-12 01:33:24 +00:00

moe_v055

fix(build): 回退qwen3_6_scripts+ex_engine到26e6cb40(能得分版本)

2026-08-12 01:33:24 +00:00

common_fused_moe_base.h

ref(EX): import upstream ILU kernels + xllm MoE CUDA sources into ex_engine

2026-08-10 03:59:45 +00:00

common_fused_moe.h

ref(EX): import upstream ILU kernels + xllm MoE CUDA sources into ex_engine

2026-08-10 03:59:45 +00:00

common_moe_fused_topk.cpp

ref(EX): import upstream ILU kernels + xllm MoE CUDA sources into ex_engine

2026-08-10 03:59:45 +00:00

common_moe_fused_topk.h

ref(EX): import upstream ILU kernels + xllm MoE CUDA sources into ex_engine

2026-08-10 03:59:45 +00:00

ex_registry.c

feat(EX): Algorithm Factor Replacement Engine — dlopen-based CUDA kernel dispatch

2026-08-10 02:25:23 +00:00

factor_gdn_chunk_fwd.cu.ref

refactor(EX): upstream-aligned kernels + FlashQLA GDN backend

2026-08-10 02:55:58 +00:00

factor_gdn_flashqla.py

refactor(EX): upstream-aligned kernels + FlashQLA GDN backend

2026-08-10 02:55:58 +00:00

factor_moe_fused_gemm.cu

feat(EX): Algorithm Factor Replacement Engine — dlopen-based CUDA kernel dispatch

2026-08-10 02:25:23 +00:00

factor_moe_topk_softmax.cu

refactor(EX): upstream-aligned kernels + FlashQLA GDN backend

2026-08-10 02:55:58 +00:00

ilu_CMakeLists.txt

perf: native ixformer decode (v1 ≤32K, v2 >32K) + flash_attn_varlen prefill

2026-08-13 07:04:21 +00:00

ilu_kernel_activation.cpp

ref(EX): import upstream ILU kernels + xllm MoE CUDA sources into ex_engine

2026-08-10 03:59:45 +00:00

ilu_kernel_attention.cpp

ref(EX): import upstream ILU kernels + xllm MoE CUDA sources into ex_engine

2026-08-10 03:59:45 +00:00

ilu_kernel_fused_moe.cpp

ref(EX): import upstream ILU kernels + xllm MoE CUDA sources into ex_engine

2026-08-10 03:59:45 +00:00

ilu_kernel_group_gemm.cpp

ref(EX): import upstream ILU kernels + xllm MoE CUDA sources into ex_engine

2026-08-10 03:59:45 +00:00

ilu_kernel_matmul.cpp

ref(EX): import upstream ILU kernels + xllm MoE CUDA sources into ex_engine

2026-08-10 03:59:45 +00:00

ilu_kernel_norm.cpp

ref(EX): import upstream ILU kernels + xllm MoE CUDA sources into ex_engine

2026-08-10 03:59:45 +00:00

ilu_kernel_rope.cpp

ref(EX): import upstream ILU kernels + xllm MoE CUDA sources into ex_engine

2026-08-10 03:59:45 +00:00

ilu_layer_attention.cpp

ref(EX): import upstream ILU kernels + xllm MoE CUDA sources into ex_engine

2026-08-10 03:59:45 +00:00

ilu_layer_attention.h

fix(build): 回退qwen3_6_scripts+ex_engine到26e6cb40(能得分版本)

2026-08-12 01:33:24 +00:00

ilu_layer_fused_moe.cpp

ref(EX): import upstream ILU kernels + xllm MoE CUDA sources into ex_engine

2026-08-10 03:59:45 +00:00

ilu_layer_fused_moe.h

fix(build): 回退qwen3_6_scripts+ex_engine到26e6cb40(能得分版本)

2026-08-12 01:33:24 +00:00

ilu_layers_CMakeLists.txt

perf: native ixformer decode (v1 ≤32K, v2 >32K) + flash_attn_varlen prefill

2026-08-13 07:04:21 +00:00

ix_full_bridge_v2.cpp

fix(build): match project_7 proven Dockerfile — remove ENV lines + .dockerignore

2026-08-13 02:48:28 +00:00

ix_full_bridge.cpp

refactor(bridge): rewrite ix_full_bridge.cpp for actual base image symbols

2026-08-10 06:34:54 +00:00

ix_moe_bridge.cpp

fix(build): 回退qwen3_6_scripts+ex_engine到26e6cb40(能得分版本)

2026-08-12 01:33:24 +00:00

moe_topk_softmax_v3.cu

fix(MoE): rewrite topk kernel — 1 block/row, shared mem, warp-agnostic

2026-08-10 08:03:08 +00:00