dylanyunlong
0 Followers · 0 Following
  • Joined on 2026-08-06
Repositories
6
 Projects Public Activity Starred Repositories
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-14 09:40:33 +08:00
1d9b620416 fix(crash): disable corex_gdn_chunk_recurrent — pybind signature mismatch
716034bdd0 fix(OOM): lower blocks cap 5000→3000 — flash_attn needs ~4GB temp buffer
Compare 2 commits »
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-14 09:26:17 +08:00
c8a982c4e8 feat: ix_full_bridge.so — dlopen bridge for ixformer::infer C++ API
872be0effa fix(build): strip \r\n from all .py files — CRLF breaks patch_ops.sh text matching
Compare 2 commits »
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-14 09:00:59 +08:00
aa4b4992d1 fix(build): hardcode blocks cap 5000 in .py — remove yaml env var
456380eed0 fix(OOM): cap GPU blocks at 5000 via BI100_MAX_GPU_BLOCKS env var
c6aa1b9c62 fix(P0): protocol.py extra=allow — recover 180 rejected replay requests
20aac5b212 fix(OOM): return zeros during profiling — skip both flash_attn AND Q-tiling
048302bd4a Revert "fix: remove --enable-chunked-prefill — conflicts with small max_num_batched_tokens"
Compare 6 commits »
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-14 00:14:21 +08:00
e31bd69779 fix(OOM): max-num-batched-tokens 4096→256 — profiling Q-tiling also OOMs
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-13 23:41:08 +08:00
2717bafc30 fix(OOM): skip flash_attn_varlen during profiling
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-13 23:10:01 +08:00
aebc660a10 revert: restore to 8c8c0286 (last confirmed build success)
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-13 22:46:06 +08:00
bed1fc4d54 fix(OOM): skip flash_attn_varlen during profiling — use Q-tiling fallback
4518a39d12 fix: restore LD_PRELOAD libcccl_allocator.so — verified compiles in docker build
Compare 2 commits »
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-13 21:59:17 +08:00
8c8c0286c9 Revert "fix(build): revert patch_ops.sh to a3c45d3b — remove cccl_preload + corex extension compile steps"
14d1725cdd fix(build): revert patch_ops.sh to a3c45d3b — remove cccl_preload + corex extension compile steps
Compare 2 commits »
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-13 21:40:49 +08:00
09d92dce5d fix(build): remove 1106 extra CCCL headers — keep only 288 needed by allocator
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-13 21:35:01 +08:00
5ec60dc574 fix(build): remove --num-gpu-blocks-override — may fail platform yaml validation
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-13 21:26:19 +08:00
71644e1530 fix: num-gpu-blocks-override 4000→5500 — 4000 blocks can't fit a 70K prompt
cf7824313f fix(OOM): add --num-gpu-blocks-override=4000 to skip profiling
8d2f30f065 fix(critical): remove LD_PRELOAD libcccl_allocator.so — crashes service if .so missing
Compare 3 commits »
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-13 20:45:01 +08:00
451bdc8204 fix: lower gpu-memory-utilization to 0.85 — prevent profiling OOM
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-13 19:37:40 +08:00
6092edebde fix: import torch.utils.cpp_extension explicitly in build script
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-13 19:36:35 +08:00
f6cf9d662e fix(CCCL): split compilation to isolate CCCL headers from torch/corex
05706f0d60 fix(build): use block-level CUB only — device-level API conflicts with corex CUDA 10.2
Compare 2 commits »
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-13 19:22:52 +08:00
daa8067080 fix: remove expandable_segments:True from PYTORCH_CUDA_ALLOC_CONF
4c365b8c03 feat(CCCL): device-level CUB algorithms for MoE dispatch
Compare 2 commits »
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-13 18:50:09 +08:00
7ba97f7977 fix: relax .dockerignore — remove *.txt *.md *.json wildcards that may break build
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-13 18:38:47 +08:00
45161610f0 fix: thread_local reentrant guard — prevent cudaMalloc infinite recursion
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-13 18:36:54 +08:00
3ce5bff10f fix: use cccl_preload::cub namespace — CUB_WRAPPED_NAMESPACE requires it
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-13 18:35:39 +08:00
c1e23615b5 fix: remove CUB_WRAPPED_NAMESPACE and _CCCL_COMPILER_GCC from build flags
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-13 18:33:20 +08:00
32325f9624 build: wire CCCL preload into competition pipeline