dylanyunlon e9eaad0592 perf: increase prefix attention tile budget 96MB→256MB ...

CCCL dispatch_transform.cuh spread_out_items_per_thread pattern:
reduce tile count = reduce Python loop iterations = faster prefill.

At 256K context with q_len=4096: old 96MB budget → 219 KV tokens/tile
→ ~1200 tiles per layer → 16 min per chunk. New 256MB budget →
~580 KV tokens/tile → ~450 tiles per layer → ~6 min per chunk.

BI-V100 has 32 GB HBM; 256 MB temporary tensor is safe.

2026-08-07 04:44:18 +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

[DEPLOY] Complete submission: baseline + all optimizations

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

_custom_ops.py

[CRITICAL/deploy] fix 3 deployment gaps found from docker crash log

2026-08-06 07:01:14 +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

[DEPLOY] Complete submission: baseline + all optimizations

2026-07-30 16:06:20 +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

[ENGINE] model_runner.py: CCCL CachingDeviceAllocator pattern — reduce CUDA graph capture from 1028→19 sizes

2026-08-07 01:59:18 +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

[CRITICAL/deploy] fix 3 deployment gaps found from docker crash log

2026-08-06 07:01:14 +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

[DEPLOY] Complete submission: baseline + all optimizations

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

qwen3_5.py

[FIX] qwen3_5.py: replace solve_triangular with manual forward substitution

2026-08-06 03:02:29 +00:00

qwen3coder_tool_parser.py

[DEPLOY] Complete submission: baseline + all optimizations

2026-07-30 16:06:20 +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

[DEPLOY] Complete submission: baseline + all optimizations

2026-07-30 16:06:20 +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

feat(verify): expand functional test suite from 21 to 51 test cases

2026-08-07 02:01:23 +00:00

xformers.py

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

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