dylanyunlong
0 Followers · 0 Following
  • Joined on 2026-08-06
Repositories
10
 Projects Public Activity Starred Repositories
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-10 10:44:56 +08:00
b4e055e9a9 feat(enginex): CCCL-style algorithm factor replacement engine — 18 operator dispatch system
b75965d4ea fix(EX): corex ivcore10 build flags + deploy pipeline + topk kernel cleanup
Compare 2 commits »
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-10 10:26:28 +08:00
fcfb764560 feat(EX): Algorithm Factor Replacement Engine — dlopen-based CUDA kernel dispatch
121432f8e9 doc: system design — architecture, file map, data flow, build pipeline
Compare 2 commits »
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-10 09:43:11 +08:00
c077736968 feat(SM70): wire up FlashQLA GDN kernel dispatch in prefill path
47958c4ed2 fix(build): add ninja dependency — required for CUDA kernel compilation
Compare 2 commits »
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-10 09:09:33 +08:00
20cd2d8904 build(SM70): precompile GDN CUDA kernel to .so during docker build
8cf73ad39c feat(SM70): add 1Cat-vLLM FlashQLA fused GDN CUDA kernel for BI-V100
3d5f75fefd fix(d05): remove image_url stripping — model IS multimodal
83d633798f fix(overflow): chunk_size 64→16 — CCCL counter overflow prevention
0a697f5871 arch(scan): dispatch_scan.cuh Phase 1/Phase 2 separation in GDN chunk loop
Compare 5 commits »
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-09 09:11:43 +08:00
a7537ebee0 doc: add functional test FAIL root cause analysis to CODEPATH_MAP
e87470733d accel(ixformer): wire BI-V100 hardware primitives into GDN + MoE compute paths
5cd2780320 fix(CRITICAL): CCCL overflow guard — clamp before cumsum + max-num-seqs=2
68876acd1b doc: BI-V100 hardware probe raw data (SSH Aug 8 2026)
6b8965a667 accel(ixformer): add BI-V100 hardware op wrappers + silence corex warnings
Compare 7 commits »
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-08 20:31:00 +08:00
c1065aaf2c fix(build): add .dockerignore + safe probe — fix docker build failure
dbfe20fd1c arch(probe): add build-time CoreX API discovery — stop guessing interfaces
ee09550263 arch(CoreX): CCCL env_dispatch — try native fused kernels, fallback PyTorch
fb2ddb843e fix(patch_ops): correct contradictory deploy log messages
Compare 4 commits »
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-08 19:04:10 +08:00
80fa1fe781 arch(CRITICAL): match Sub168 proven engine config exactly
1fed1bc051 fix: add --max-seq-len-to-capture 32768, fix patch_ops.sh contradictory comments
Compare 2 commits »
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-08 18:49:45 +08:00
d221383fc0 doc(prd): complete base engine migration checklist
6bed911e04 fix(patch_ops): add pip install transformers==4.55.3 from base engine
e0fe46a46f arch(CRITICAL): deploy ALL base engine patches — paged_attn, xformers, sequence, scheduler
Compare 3 commits »
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-08 16:11:51 +08:00
abd3d5640a arch(CRITICAL): replace custom qwen3_5.py with base original (1369 lines)
4daa30a267 fix(CRITICAL): CCCL kernel_segmented_scan — kill nan_to_num, add corex_gdn/corex_moe dispatch
c077f7fd40 doc(prd): add block_reduce + exception mapping records
e832687893 arch(qwen3_5): dispatch_segmented_sort three-way dispatch — try native CoreX before PyTorch fallback
d44ec4d8db perf(qwen3_5): CCCL block_reduce_warp_reductions → reduce DeltaNet loop iterations
Compare 9 commits »
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-08 15:44:25 +08:00
85f3240c98 doc(prd): create PRD with CCCL→base mapping table and competition strategy
ef540b6f9f fix(serving): CCCL completion_mechanism — remove NaN-era max_tokens cap
68be2ff856 fix(dispatch): radix_sort-inspired size-dispatch — disable thinking for small max_tokens, clamp oversized max_tokens
e37b4d283b env(yaml): CCCL buddy_allocator pattern — PYTORCH_CUDA_ALLOC_CONF + OMP_NUM_THREADS
b271210af4 fix(critical): allow n=2 to match Sub168 — max_num_seqs=2 in yaml
Compare 5 commits »
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-08 15:19:31 +08:00
e86b7eacdd refactor(tool_parser): CCCL agent_radix_sort_downsweep union TempStorage — phased state
5a3831e977 refactor(api): CCCL tuning_adjacent_difference policy_selector — dynamic error handling
8be95e422d perf(protocol): CCCL agent_for consume_tile<IsFullTile> — fast path for standard messages
2ea7a19f73 refactor(serving): CCCL dispatch_rle streaming_context pattern — unify per-choice state
cafd34fe4a fix(CRITICAL): REVERT to serving-only patches — Sub168 proves CoreX native model is correct
Compare 8 commits »
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-07 18:30:52 +08:00
2680d62ec8 fix(critical): match Sub168 config exactly + disable risky numerical patch
1ba0dd3966 arch(cccl): match Sub168 proven config + bench.py timeout pattern
3342d18bcc fix(critical): remove pip install transformers — was breaking corex kernel loading
4e04674283 fix(cccl): robust fallback — inject module-level safety if regex misses
7153029974 perf(cccl): thread_reduce fast-path — cap tool_call max_tokens to 2048
Compare 7 commits »
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-07 17:30:10 +08:00
4eea584c9d fix(deltanet): systematic NaN elimination via CCCL overflow_cast pattern
a20e8614a4 fix(critical): stop replacing base image compute files — use corex native kernels
2102146c01 refactor(moe): translate thrust mode.cu pipeline — unique_consecutive replaces manual boundary detect
4d9165fa30 arch(tuning): translate CCCL tuning_transform_tile.cuh — derive chunk sizes from hardware
f140825a56 arch(moe): translate CCCL sync_handler.cuh — register-at-init, resolve-on-first-call
Compare 20 commits »
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-07 16:38:28 +08:00
840fe923cc fix(critical): DeltaNet NaN 99.98% — clamp gate logits before exp to prevent overflow
57a2216143 fix: max_num_seqs=2 for n=2 support + remove protocol n clamp
Compare 2 commits »
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-07 16:14:42 +08:00
ca3848dae1 fix(critical): 3 fixes from sub508 diagnosis — n>1 crash guard + thinking format + content fallback
05c775ca11 docs: complete sub508 diagnosis — 19 cascade failures from n=2, full test matrix
994c6575af fix(critical): clamp n>1 to 1 in protocol — prevent t2_n_2 engine crash cascade
c241764506 fix(critical): wrap get_scheduler_config in try-catch — prevent n=2 engine crash cascade
054544b7a8 docs: sub509 diagnosis — d03 FAIL root cause + performance gap analysis
Compare 5 commits »
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-07 15:47:13 +08:00
bc64366d9e comp 168 submit
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-07 15:46:18 +08:00
e0344b1730 fix(critical): disable thinking for tool_call requests — fixes d03_tool_call FAIL
812c374f7a fix(critical): sync baseline.muh max_model_len=100000 gpu_mem=0.90 — match computility-run.yaml
19879ccaae docs: pipeline ground truth — CCCL parity audit, SM100 benchmark data, injection status
611c491b0f audit: CCCL vs muh parity check — scale_mem_bound 11/11 PASS, SM100 benchmark values extracted
95d03147e7 fix(deploy): reduce max-model-len 131072→100000, remove chunked-prefill, single-seq — prevent OOM crash
Compare 7 commits »
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-07 15:10:41 +08:00
cbd1f08a3e fix(protocol): clamp max_tokens to available context — fix t3_max_tokens_max HTTP 400
16f0b30d2e fix(critical): 3 bugs causing 0.0 score — OOM crash + thinking param + multimodal
539fe7745b fix(protocol): accept max_completion_tokens + thinking + tool_calls messages
dd077e1272 [ENGINE] CCCL dispatch_common.cuh use_default pattern: accept max_completion_tokens + thinking + content=None
8002900af0 [ENGINE] CCCL kernel_scan.cuh dual-algorithm dispatch: v1 (online norm) vs v2 (deferred norm)
Compare 6 commits »
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-07 14:22:01 +08:00
b86a121d6d fix(critical): patch_ops.sh cwd bug — all cp/deploy ./paths resolved against Dockerfile WORKDIR=/workspace/ instead of script dir
dylanyunlong pushed to main at dylanyunlong/project_6 2026-08-07 13:33:14 +08:00
36e67c00b0 fix(critical): deploy patches to BOTH lib and lib64 vllm paths
e9eaad0592 perf: increase prefix attention tile budget 96MB→256MB
025059d78e fix(critical): raise decode threshold to prevent service crash
Compare 3 commits »