Skip to content
GitLab
Explore
Sign in
铁森林
2026abc
Repository
Branches
Overview
Active
Stale
All
test
9e55ad61
·
Delete 333
·
Aug 06, 2026
ablation/L9-full
700bca1e
·
ablation L9: full a193 default profile (PRA26_QWEN35_TP2_PROFILE=1, no...
·
Aug 06, 2026
ablation/L8-w8a8-attnout-a55-a56
5b1eedcf
·
ablation L8: W8A8 attn-output + concurrent decode INT8 fusions A55/A56 on top...
·
Aug 05, 2026
ablation/L7-w8a8-gdn-qkv
06a09f9c
·
ablation L7: W8A8 GDN QKVZ + full QKV all layers on top of L6 (non-equivalent,...
·
Aug 05, 2026
ablation/L6-w8a8-mlp
871de843
·
ablation L6: W8A8 dense MLP all 64 layers on top of L5 (non-equivalent, may change token selection)
·
Aug 05, 2026
ablation/L5-lmhead-w8r4
c353705d
·
ablation L5: LM-head W8+R4 quant + rocblas w8a8 + tp local greedy on top of L4...
·
Aug 05, 2026
ablation/L4-tp2-comm
f68c57d1
·
ablation L4: TP2 comm fusion - gfx936 custom allreduce + AR+GemmaRMSNorm(mlp)...
·
Aug 05, 2026
ablation/L3-gdn-fusion
2e16e4ee
·
ablation L3: GDN QKVZ INT32 postproc fusion on top of L2 (math-equivalent)
·
Aug 05, 2026
ablation/L2-paged-fa
8e752987
·
ablation L2: Paged-KV FlashAttention direct read (CUSTOM backend, attention only)
·
Aug 05, 2026
ablation/L1-tunableop
6a3a20af
·
ablation L1: TunableOp GEMM selection (PYTORCH_TUNABLEOP_ENABLED=1)
·
Aug 05, 2026
ablation/L0-baseline
b355a3a0
·
ablation L0: baseline config (all PRA26 optimizations off, TRITON_ATTN backend)
·
Aug 05, 2026
codex/f12b-causal-conv-signature-20260804
72c0f50a
·
Warm Qwen3.5 causal-conv production signature
·
Aug 04, 2026
codex/f12c-gdn-complete-signature-20260804
b887e572
·
Warm Qwen3.5 GDN production signature
·
Aug 04, 2026
codex/a01c-dp2-least-active-20260803
55fe8f08
·
Optimize C8 replica routing with least-active dispatch
·
Aug 03, 2026
exp/a53-tp2-w8a8-20260803
5f9d490f
·
Add doc 08: multi-concurrency diagnosis (chunked-prefill + 65us latency, bt8192 worse)
·
Aug 03, 2026
submit-a193-tp2-concurrent
148e405e
·
Submit TP2 concurrent inference optimizations
·
Aug 02, 2026
submit-311-tp2-w8-block-parallel
477ec9bf
·
perf: parallelize ordered W8 down projection blocks
·
Jul 29, 2026
exp/a50-tp2-vendor-prefill-12h2kv-20260727
1565ce0c
·
Fix TP2 vendor paged KV prefill shape
·
Jul 27, 2026
exp/dense-w8a8-fullqkv-a50-20260720
ef403276
·
Explore full-attention QKV W8 path
·
Jul 20, 2026
exp/dense-w8a8-gdn-a49-20260720
ddd185c4
·
Explore GDN dual-weight W8 runtime dispatch
·
Jul 20, 2026
Prev
1
2
3
Next