root
|
41b51382fd
|
prebuilt: corex_gdn_chunk_recurrent.so + remove __pycache__ from tracking
|
2026-08-14 02:12:32 +00:00 |
|
muh-bot
|
5aba296eba
|
[muh] gen_patch: expand VLLM_INJECTION_POINTS to full real injection surface
- Replace DEAD csrc/*.cu targets with 11 confirmed Python/Triton injection points
- Add paged_attn.py: _PARTITION_SIZE, use_v1 (V1/V2 dispatch threshold)
- Add computility-run.yaml: max-num-seqs, max-num-batched-tokens, gpu-mem-utilization
- Preserve Triton autotune injection: flash_attn BLOCK_M/N, prefix_prefill BLOCK/NUM_WARPS
- Fix PARTITION_SIZE semantic: tile size (threads*items), not items_per_thread alone
- Document CCCL parallels for each injection point
- Validated: gen_patch --dry-run produces patch (reduce -> paged_attn.py)
- Validated: test_smem_safety.py 191/191 all safe
- Validated: scale_mem_bound CCCL parity 14/14 pass
|
2026-08-06 04:01:40 +00:00 |
|
muh-bot
|
5e618bf480
|
[housekeeping] add enginex_base/ to gitignore — 解压后的基础引擎仅供本地分析
基础引擎结构 (排除 pkgs 预编译包):
顶层: Dockerfile, computility-run.yaml, paged_attn.py, prefix_prefill.py, attention.py
qwen3_6_scripts/: 我们的全部适配代码 (模型定义, attention, API, parsers, patches)
vllm/: vllm 0.6.3 BI-V100 适配版 (985 files)
_custom_ops.py: ixformer 绑定层 (V1 decode, silu_and_mul, rms_norm...)
attention/selector.py: backend 选择 (BI-V100 → XFORMERS → _run_sdpa_fallback)
model_executor/models/: 200+ 模型定义 (无 qwen3_5)
关键发现:
1. 基础引擎 Dockerfile 只有 4 行: FROM + COPY + RUN patch_ops.sh
2. vllm/_custom_ops.py 硬编码 SMEM=32KB (已在我们版本修为 49152)
3. attention selector: BI-V100 → XFORMERS (not FlashAttn, SM<80)
4. Prefill 路径: xformers._run_sdpa_fallback (PyTorch matmul+softmax)
不是 paged_attn.forward_prefix (那是 context attention for cached tokens)
CCCL source read: dispatch_batch_memcpy.cuh (large_buffer vs small_buffer paths)
KV block size = 16×4×256×2 = 32KB → large_buffer path in CCCL
|
2026-08-05 07:17:37 +00:00 |
|
dylanyunlon
|
ef6abf3dc7
|
[DEPLOY] Complete submission: baseline + all optimizations
Adds ALL files needed for Dockerfile build:
- qwen3_6_scripts/ (baseline patches + our optimizations)
- vllm/ (full vllm package)
- paged_attention_v2_pytorch.py (V2 with single-bmm optimization)
- Dockerfile + computility-run.yaml
Our optimizations vs baseline:
1. paged_attn.py: pre-gathered context KV (eliminates 194 gather calls),
Triton try/fallback, V2 heuristic, threshold 32K→64K
2. paged_attention_v2_pytorch.py: fills NotImplementedError,
single-bmm Phase 1 (195 launches → 3)
3. patch_enable_triton.py: HAS_TRITON=True with safety fallback
4. patch_triton_tuning.py: BLOCK=64, NUM_WARPS=4 for BI-V100
5. computility-run.yaml: gpu-memory-utilization 0.9→0.95,
max-num-batched-tokens 8192→16384
This repo can now be submitted to dev.modelhub.org.cn as-is.
|
2026-07-30 16:06:20 +00:00 |
|
dylanyunlon
|
9b21a13119
|
[MUH] Bootstrap muh toolchain — extract/parse/gen_yaml/gen_patch + baseline.muh
Pipeline:
1. extract.py: Parses all 26 CCCL tuning_*.cuh → 26 YAML schemas in muh/schema/
2. parse.py: .muh file parser with extends-inheritance + schema validation
3. gen_yaml.py: .muh → computility-run.yaml (verified: matches competition reference)
4. gen_patch.py: .muh → vllm kernel unified diff patches (6 algorithm mappings)
5. baseline.muh: Competition reference config, all tuning values pending BI-V100 benchmarks
Schemas extracted:
26 algorithms, 8-19 params each, SM75/80/90/100 reference tunings
Priority mapping: reduce→attention, topk→sampling, scan→paged_attention,
transform→activations, batch_memcpy→KV_cache, for→RoPE
Tested: extract→parse→validate→gen_yaml→gen_patch full pipeline passes
|
2026-07-30 10:39:06 +00:00 |
|