project6 43ede018a1 perf(moe): translate CCCL smem_resource_raw.cuh — buffer reuse for MoE output ...

smem_resource_raw.cuh entire design (180 lines):
  Manages shared memory as multi-stage pipeline resources.
  Core idea: one memory region, multiple stages, barrier-synchronized.
  - mStageCount stages share the same SMEM base pointer
  - data() returns mPtrBase + mStageCurrent * mStride (stage rotation)
  - incrementStage() rotates, parity flips on wraparound
  - release/acquire protocol for producer-consumer sync
  Key insight: allocate once, reuse forever via stage rotation + zeroing.

Translation to MoE _pure_pytorch_experts:
  Previous: torch.zeros_like(hidden_states) every call — GPU malloc + memset.
  Now: class-level _moe_out_buf, resized only when shape changes, .zero_()
  in-place (memset only, no malloc). On BI-V100 without async allocator,
  this eliminates a synchronous cudaMalloc per MoE layer per forward pass.
  With 28 MoE layers × 2 calls/step (prefill+decode), that is 56 fewer
  allocations per step.

CCCL source: cub/cub/detail/warpspeed/resource/smem_resource_raw.cuh
Maps to: qwen3_6_scripts/qwen3_5.py (Qwen3_5MoeSparseBlock)

2026-08-07 09:13:32 +00:00

..

qwen3_5

[DEPLOY] Complete submission: baseline + all optimizations

2026-07-30 16:06:20 +00:00

qwen3_5_moe

[DEPLOY] Complete submission: baseline + all optimizations

2026-07-30 16:06:20 +00:00

reasoning

fix(critical): 3 fixes from sub508 diagnosis — n>1 crash guard + thinking format + content fallback

2026-08-07 07:55:04 +00:00

_custom_ops.py

fix(critical): CCCL policy_selector degradation for MoE — PyTorch fallback for topk_softmax

2026-08-07 08:56:50 +00:00

api_server.py

[DEPLOY] Complete submission: baseline + all optimizations

2026-07-30 16:06:20 +00:00

arg_utils.py

[ARCH] Eliminate AST patch scripts — full file replacements only

2026-08-05 08:24:43 +00:00

chat_utils.py

fix(critical): 3 bugs causing 0.0 score — OOM crash + thinking param + multimodal

2026-08-07 07:05:40 +00:00

cli_args.py

[DEPLOY] Complete submission: baseline + all optimizations

2026-07-30 16:06:20 +00:00

logits_processor.py

[ARCH] Eliminate AST patch scripts — full file replacements only

2026-08-05 08:24:43 +00:00

mamba_cache.py

[fix/correctness] mamba_cache: safe swap via clone, not in-place fancy indexing

2026-08-06 06:33:26 +00:00

model_runner.py

fix(engine): CCCL overflow_cast + checked_allocator patterns for NaN/OOM

2026-08-07 08:56:36 +00:00

paged_attention_v2_pytorch.py

[ENGINE] CCCL SmemResource pattern: pre-allocate staging buffers in V2 paged attention

2026-08-07 04:37:44 +00:00

paged_attn.py

perf: increase prefix attention tile budget 96MB→256MB

2026-08-07 04:44:18 +00:00

patch_ops.sh

fix(critical): patch_ops.sh cwd bug — all cp/deploy ./paths resolved against Dockerfile WORKDIR=/workspace/ instead of script dir

2026-08-07 06:20:02 +00:00

patch_transformers_qwen3_5.py

[DEPLOY] Complete submission: baseline + all optimizations

2026-07-30 16:06:20 +00:00

prefix_prefill.py

fix(critical): v2 kernel softmax normalization was commented out — outputs were wrong

2026-08-07 02:46:46 +00:00

protocol.py

fix: max_num_seqs=2 for n=2 support + remove protocol n clamp

2026-08-07 08:18:41 +00:00

qwen3_5.py

perf(moe): translate CCCL smem_resource_raw.cuh — buffer reuse for MoE output

2026-08-07 09:13:32 +00:00

qwen3coder_tool_parser.py

fix(critical): disable thinking for tool_call requests — fixes d03_tool_call FAIL

2026-08-07 07:45:28 +00:00

registry.py

[ARCH] Eliminate 2 more patch scripts — registry.py + tool_parsers __init__.py

2026-08-05 08:36:52 +00:00

sampler.py

[BASE] qwen3_6_scripts/sampler.py: CCCL topk unsorted output optimization

2026-08-06 02:55:51 +00:00

scheduler.py

[DEPLOY] Complete submission: baseline + all optimizations

2026-07-30 16:06:20 +00:00

sequence.py

[DEPLOY] Complete submission: baseline + all optimizations

2026-07-30 16:06:20 +00:00

serving_chat.py

fix(thinking): default enable_thinking=True for t1a/t1c PASS

2026-08-07 09:01:57 +00:00

tool_parsers_init.py

[ARCH] Eliminate 2 more patch scripts — registry.py + tool_parsers __init__.py

2026-08-05 08:36:52 +00:00

verify_functional.py

fix(verify): remove duplicate TC-22~30 test definitions — CCCL test_then.cu audit

2026-08-07 06:36:12 +00:00

xformers.py

feat(xformers): replace Q-only tiling with Q+KV tiling + online softmax

2026-08-07 01:54:52 +00:00