Commit Graph

Select branches

Hide Pull Requests

cch/megamoe-rc-v0.1

Mono Color

• ebf11be355 fix(moe): gate fp8 megamoe on deepgemm capability cch/megamoe-rc-v0.1 LuminolT 2026-07-08 12:05:46 +08:00
• 89ba17ad05 feat(moe): add fp8 megamoe fused path and nvfp4 layout guards LuminolT 2026-07-08 11:42:51 +08:00
• eb7f44a8ee fix(moe): pass megamoe fp4 weights as int8 to deepgemm LuminolT 2026-07-07 14:31:03 +08:00
• f482a06a51 test(moe): exercise megamoe a2a handler without dummy shortcut LuminolT 2026-07-06 14:16:01 +08:00
• 0d3bf6b5c2 fix(moe): import capture mode from cuda graph runner LuminolT 2026-07-06 14:13:55 +08:00
• c00630088c fix(moe): infer megamoe fp4 weight scale group LuminolT 2026-07-06 10:46:18 +08:00
• 93e3840578 test(moe): add GLM MegaMoE validation coverage LuminolT 2026-07-06 10:44:05 +08:00
• 57bf729af9 feat(glm): route GLM MoE through megamoe backend LuminolT 2026-07-06 10:39:23 +08:00
• c13d4556ff feat(moe): add megamoe forward and weight layout LuminolT 2026-07-06 10:37:21 +08:00
• beb2162028 feat(moe): add megamoe a2a backend plumbing LuminolT 2026-07-06 10:26:43 +08:00
• 6f96a1c269 Enable flashinfer TRTLLM for FP8 EAGLE MoE wxiwnd 2026-06-30 00:05:16 +08:00
• 2de822661c Avoid decode-step syncs in min-new-token penalties laoyao0822 2026-06-27 20:37:05 +08:00
• 132561a151 Preserve speculative finish correctness across reasoning and stop strings laoyao0822 2026-06-27 20:36:52 +08:00
• f3cf95c0f1 Release decode transfer state on abort paths laoyao0822 2026-06-27 01:57:31 +08:00
• ac47eb61c9 Reuse draft MTP indexer topk on spec v1 EAGLE laoyao0822 2026-06-28 00:57:31 +08:00
• df3cc9abf8 Reuse draft MTP indexer topk only when model config allows laoyao0822 2026-06-28 00:41:04 +08:00
• 648a33ab30 Remove NSA spec-v2 graph metadata host syncs laoyao0822 2026-06-28 02:29:23 +08:00
• 4414db594c Reduce spec-v2 overlap copy overhead laoyao0822 2026-06-27 09:07:37 +08:00
• c6b99f6060 Stabilize spec-v2 draft graph metadata laoyao0822 2026-06-27 02:35:09 +08:00
• a27114d9dc Prevent spec-v2 decode warmup races laoyao0822 2026-06-27 02:12:37 +08:00
• 014948c5d7 CP shared-KV: global-position-rotated zigzag owner-lane balancing (default on) leavelet 2026-06-28 14:09:58 +00:00
• ce0e486327 Gate scheduler metrics on attn_cp_rank==0 too, not just attn_tp_rank==0 leavelet 2026-06-25 06:27:52 +00:00
• 01d6ceff8f fix(nsa): guard cp split to extend batches wxiwnd 2026-06-26 22:50:35 +08:00
• c739ea667d Keep CP HiCache backup state replicated on radix nodes laoyao0822 2026-06-26 08:17:41 +08:00
• 0da83747cf CUDA graph: opt-in per-buffer + per-shape decode CG memory logging leavelet 2026-06-24 16:28:15 +00:00
• 144ae7acb1 Regression test: CP HiCache L3 reload same-tick admit->reserve race leavelet 2026-06-24 14:52:30 +00:00
• 8cd46c18b3 Fix CP HiCache L3 reload duplicate-reservation crash: recheck clean-attach at reserve leavelet 2026-06-24 14:48:37 +00:00
• fdc23f7d61 CP HiCache: redesign owner-lane eviction planner as a lazy-re-evaluation heap (O(n log n)) leavelet 2026-06-24 01:35:37 +00:00
• 2e15cb257a bench: CP owner-lane eviction planner micro-benchmark (validates the de-sync fix) leavelet 2026-06-24 01:25:18 +00:00
• 4e9b4d05c0 CP HiCache: kill the O(victims*candidates) GPU-sync storm in the owner-lane eviction planner leavelet 2026-06-24 01:13:59 +00:00
• ff1804f5cf CP HiCache perf: loud one-time warning when PLACEMENT_ASSERT is on in a multi-rank run leavelet 2026-06-23 18:11:45 +00:00
• ea4ce31f95 CP HiCache L3 3.4: cold-start (restart durability) — load|clear via SGLANG_CP_L3_COLD_START leavelet 2026-06-23 18:09:37 +00:00
• d6aabace90 CP HiCache: fix L3-reload partial-prefix node split crash (well-form the node) leavelet 2026-06-23 17:01:19 +00:00
• 47c24fecf2 CP HiCache metrics: fix tp_rank AttributeError under CP + label by cp_rank leavelet 2026-06-23 16:06:51 +00:00
• dbc5ebbaa0 CP HiCache: multi-slab host KV cache to respect cudaHostRegister per-call ceiling leavelet 2026-06-23 15:33:10 +00:00
• 0025b0e819 fix: count huge pages leavelet 2026-06-23 10:29:52 +00:00
• 6114c3e3e5 HiCache host pool: check hugepages, not regular RAM, for the hugetlbfs-backed (CP shared-L2) pool leavelet 2026-06-23 10:17:33 +00:00
• 9dd83bcb88 CP HiCache metrics: cover pure-L2 (no L3), lazily create the collector when L2 pooling is on leavelet 2026-06-23 09:37:11 +00:00
• 5163ba7bb9 CP HiCache: expose L2 + L3 usage metrics (the mainline storage metrics are dead under CP) leavelet 2026-06-23 02:01:05 +00:00
• e96cfa6cbd L3 3.3: continuous-background, communication-free, page-level LRU GC (+ wire the reload touch) leavelet 2026-06-23 01:39:47 +00:00
• 864b1c808e L3 3.2: async disk->L2 reload + request hold (Model B), zero new collectives, EAGLE-correct leavelet 2026-06-23 00:09:38 +00:00
• 815da6c4e5 L3 3.1: redesign spill — 2-phase staging pipeline + proactive FIFO (fix evicted=0 starvation) leavelet 2026-06-22 22:15:48 +00:00
• f353cf4702 Revert "Fix CP HiCache writing_check deadlock at the source: rank-replicated symmetric attach" leavelet 2026-06-22 21:35:27 +00:00
• d08ee9f8d9 L3 3.1: background spill maintainer (resident->L3 copy, approach D) leavelet 2026-06-21 05:20:50 +00:00
• 2ae46013ac L3 3.0: wire CpL3Store into HiRadixCache (construct + drain + reset/clear) leavelet 2026-06-20 23:30:41 +00:00
• 81a1b72be6 L3 3.0: wire --enable-cp-l3 flag, validation, hash-from-init leavelet 2026-06-20 23:25:46 +00:00
• de1d1e0af2 L3 3.0: CpL3Store orchestrator (async spill/reload, object-granular acks) leavelet 2026-06-20 23:22:10 +00:00
• 2ea3728a43 L3 3.0: host-slab accessor (layer_page_first <-> contiguous blob) leavelet 2026-06-20 23:17:31 +00:00
• 99a695d747 L3 3.0: POSIX disk backend (CpL3DiskSlab, O_DIRECT page slots) leavelet 2026-06-20 23:15:47 +00:00
• bff965a62e L3 3.0: shared per-node LMDB metadata index (CpL3MetaStore) leavelet 2026-06-20 23:11:04 +00:00
• b8ab180f1a L3 3.0: config (disks + rank->disk mapping + PLP gate) leavelet 2026-06-20 23:08:24 +00:00
• 04c6793ca4 L3 3.0: on-disk blob format + disk-slot free-list allocator leavelet 2026-06-20 23:05:55 +00:00
• 85585fcf4b Fix CP HiCache writing_check deadlock at the source: rank-replicated symmetric attach leavelet 2026-06-22 19:42:41 +00:00
• bed4601a20 B300 CP decouple: pad-aware hidden-states gather/scatter for standard a2a under shared-KV leavelet 2026-06-21 19:06:53 +00:00
• f0c6b892b6 Release CP draft-hidden buffers after EAGLE target prefill laoyao0822 2026-06-22 01:05:49 +08:00
• 1d7149df1c Avoid remap-cache aliasing across tensor views laoyao0822 2026-06-22 01:05:41 +08:00
• c1637cfb68 B300 NextN fp4: review cleanups — keep else-slice, non-routed draft trtllm, weight_scale_2 probe leavelet 2026-06-21 17:48:06 +00:00
• 2b8593ff24 B300 NSA: default NSA prefill to flashmla under context parallelism (fix k_rope CP crash) leavelet 2026-06-21 17:33:49 +00:00
• b3a4f4174f B300 NextN fp4: run the EAGLE draft on the same trtllm MoE backend as main (elegant root fix) leavelet 2026-06-21 17:32:01 +00:00
• de726b4e7e B300 NVFP4 MoE: slice global input scales to EP-local on the non-flashinfer process_weights path leavelet 2026-06-21 17:02:43 +00:00
• 1d96197d3d B300 NextN fp4: decide keep-vs-drop from the checkpoint, not the MoE runner backend leavelet 2026-06-21 16:48:05 +00:00
• a3d0c2394b B300 NextN fp4: also honor speculative MoE runner backend in keep-decision (+ debug log) leavelet 2026-06-21 16:32:59 +00:00
• 5e670760bd B300 NVFP4: fix FlashInferFP4MoE.forward attr names to match HEAD weight-prep leavelet 2026-06-21 16:24:17 +00:00
• ae5e98cf86 B300 NVFP4 port: advance layers/quantization/utils.py to HEAD (piecemeal-gap fix) leavelet 2026-06-21 16:08:46 +00:00
• 2aca99d677 B300 NextN fp4: keep modelopt_fp4 for fp4-capable runners + generalize MTP exclude-remap leavelet 2026-06-21 16:00:23 +00:00
• c616bb7095 B300 CP decouple from DeepEP (Strategy B), stage A: relax forced moe_a2a_backend + allow tp4/CP4 leavelet 2026-06-21 15:05:45 +00:00
• 2670c6c51b B300 NVFP4 port: add get_cuda_driver_bindings to utils.common (C3 comm_fusion dep) leavelet 2026-06-21 14:53:05 +00:00
• 7225a17c58 B300 NVFP4 port P1 stage-2: C2 flashinfer_cutedsl MoE runner + C3 flashinfer comm-fusion leavelet 2026-06-21 14:50:37 +00:00
• b0baea52ef B300 NVFP4 port P1 stage-1: advance flashinfer_trtllm NVFP4 MoE + modelopt_fp4 path to upstream HEAD leavelet 2026-06-21 14:36:33 +00:00
• 9368c33b23 L2 prereqs for CP8DP2 + L3: scope HiCache consensus to the CP group + reset pooled allocator on flush leavelet 2026-06-20 21:24:54 +00:00
• 9549b268d7 Avoid allocating indexer state for shared NSA layers laoyao0822 2026-06-21 05:16:30 +08:00
• 187b700406 Keep speculative grammar traversal scalar-safe laoyao0822 2026-06-21 05:16:21 +08:00
• 9f7c193eb1 2b GC follow-on: drop dead CpSharedL2NodeMetadata.{committed, committed_payload_layers} leavelet 2026-06-20 19:36:51 +00:00
• 93fe864779 2b.3 observability: repoint lane-stats occ/used at the shared pool when pooling on leavelet 2026-06-20 19:16:44 +00:00
• 9ac42450e1 2b.3 observability: log shared-L2 pool report in lane-stats leavelet 2026-06-20 18:43:25 +00:00
• eefc678660 2b.3 fixes: restore two l2_pooling-adopt gaps surfaced by live validation leavelet 2026-06-20 18:02:53 +00:00
• 07c2b9bc57 2b.2: B1 collective-free shared-pool evict-to-fit + delete the #5 deadlock leavelet 2026-06-20 16:03:38 +00:00
• 0d87bc576f 2b dead-island GC: collapse pooled-L2 allocator to the B1 commit model leavelet 2026-06-20 15:47:45 +00:00
• 01029c8df4 2b.1b: B1 collective-free pooled-L2 write-through (default-off) leavelet 2026-06-20 13:59:39 +00:00
• 50bc923ad6 2b.1b step 1: mark_object_committed (B1 MIN-driven commit) leavelet 2026-06-20 12:51:22 +00:00
• 7850aab1a2 2b.1a: placement_digest gate on CpSharedL2PageAllocator (B1 proof obligation) leavelet 2026-06-20 12:32:25 +00:00
• 17cbfa31c9 E1 lane-stats instrumentation + 2a reserve-at-admission (default-off) leavelet 2026-06-20 12:03:14 +00:00
• 7de882b622 2b.0: adopt l2_pooling pooled-L2 mechanics (default-off) leavelet 2026-06-20 11:57:25 +00:00
• 8f1e85a992 CP HiCache: replicated-clock SLRU eviction, collective-safe (Phase 1) leavelet 2026-06-19 18:59:01 +00:00
• 8d73919d02 Gate CP HiCache write-backup on hard deficit, not the free-room watermark leavelet 2026-06-19 12:14:36 +00:00
• 96158fa110 Bound overlap prefill to one pending CP batch laoyao0822 2026-06-16 02:47:35 +08:00
• 512fe92a83 Fix CP HiCache catch_up_all_layers fallback on chunked-prefill final chunk leavelet 2026-06-18 11:54:24 +00:00
• a4ec2236ef Fix spec-decode grammar mask crash under xgrammar 0.2.1 (strict tvm-ffi binding) leavelet 2026-06-18 07:16:49 +00:00
• 40738f59cf Cache EnvField reads to avoid per-call os.getenv in the hot path leavelet 2026-06-17 20:06:55 +00:00
• 820790b97f Remove CP HiCache debug instrumentation leavelet 2026-06-17 19:59:01 +00:00
• 2212963a6c Harden OpenAI tool-call/chat-template + reasoning parsing leavelet 2026-06-17 19:58:42 +00:00
• e23168e7f5 Fix CP shared-KV bs=1 cache-hit zero-prefix corruption; spans are sole compose input leavelet 2026-06-17 18:25:22 +00:00
• b34e7cb932 Add SGLANG_CP_TRANSFER_LOG: decouple CP shared-KV transfer-partition dumps from debug flag leavelet 2026-06-17 16:29:31 +00:00
• 268bcf78da Dump: add 'fh' mode (final_hidden only, fresh prefills allowed) for cache-bust vs cache-hit leavelet 2026-06-17 16:10:41 +00:00
• d6aac79eb1 Lean tensor dump: residual trajectory only (attn_in/attn_out + q_all), drop kv/MoE leavelet 2026-06-17 15:09:39 +00:00
• 9031a294b3 Complete per-forward tensor dump: all-layer trajectory + final hidden, flush per (rid,rank) leavelet 2026-06-17 14:35:43 +00:00
• ed42a6dcbc Honor x-request-id as rid + env-gated NSA tensor dump for L1-vs-L2 reload diff leavelet 2026-06-17 14:05:54 +00:00
• e4e0784387 CP HiCache trace: content-key in fwd_hash for L1-hit vs L2-reload differential leavelet 2026-06-17 11:20:54 +00:00
• 25c2d6d606 CP HiCache trace: cross-rank loc-divergence confirm (visible_locs_hash + victim_set + release_draw) leavelet 2026-06-17 08:29:54 +00:00
• 1d54b0b7ab CP HiCache trace: reload per-page owner-check (the make-or-break invariant) leavelet 2026-06-17 04:52:17 +00:00