Commit Graph

Select branches

Hide Pull Requests

main

Mono Color

• c98a8d6648 [Bugfix] Shared host arena: do not require seal constants in fcntl (#1404) main 3 people 2026-11-06 23:36:50 +08:00
• ad339d82ed [Perf][Core] MRV2 warmup: cover every request count, mixed steps and a full chunk (#1402) 3 people 2026-11-06 23:07:00 +08:00
• 4a6e6124d9 [Bugfix][Core] Restore partial prefix hits on long hybrid prompts (27B and Flash-Next) (#1401) 3 people 2026-11-06 22:30:53 +08:00
• 171acceded [Bugfix][SM70] GDN prefill: use the prebuilt kernel when TileLang has no nvcc (#1400) 3 people 2026-11-06 21:57:50 +08:00
• ad80aaa02b [Bugfix][Frontend] Answer rejected requests without a server traceback (#1399) 3 people 2026-11-06 21:48:46 +08:00
• adafab1225 [Perf][SM70] DFlash2: keep compact target rejection with penalties (#1398) 3 people 2026-11-06 21:15:36 +08:00
• 2038fbf51b [Release] Bring the bundled V100 launchers up to the 1.5.2 profiles (#1397) 3 people 2026-11-06 16:56:50 +08:00
• c4820d05fb [Security] Drop the hard-coded API key from the DS-V4.1 serving config (#1396) 3 people 2026-11-06 16:55:30 +08:00
• 2bb8c2b9cd [Docker] Ship the CUDA 12.8 compiler in the SM70 release image (#1395) 3 people 2026-11-06 14:36:39 +08:00
• 9f970d6288 [Bugfix][SM70] Stop compiling prefill kernels for every new prompt length (#1394) 3 people 2026-11-06 09:48:19 +08:00
• d4adfda98e [DFlash2] Run the compact sampler graph for uniform verifier batches (#1393) 3 people 2026-11-06 07:42:17 +08:00
• eef9d23009 [Bugfix][SM70] Release config: Flash-Next host KV default, 27B at c8, release image (#1392) 3 people 2026-11-06 07:27:29 +08:00
• 46f3dd1b26 [Bugfix][Memory] Profile startup memory under inference mode again (#1391) 3 people 2026-11-06 06:58:58 +08:00
• 6ed2096b8e [Bugfix][Compile] Keep clones that change layout in the clone elimination pass (#1359) LeonandClaude Opus 5.5 2026-11-05 15:58:19 -07:00
• 73b03e56d0 [Bugfix][Compile] Keep returned clones that would alias graph inputs or outputs (#1358) LeonandClaude Opus 5.5 2026-11-05 15:58:09 -07:00
• 39a1a971c6 [Bugfix] Preserve streaming tool calls, errors and logprob metadata (#1317) LeonandCodex 2026-11-05 15:58:00 -07:00
• 2dd8a92685 [Bugfix] Share Qwen XML tool and answer union with qwen3_coder (#1312) Leon 2026-11-05 15:57:50 -07:00
• 3c051b52ed [Perf][SM70] Flash-Next: concurrent MTP4 verification batches (#1369) 3 people 2026-11-05 22:45:28 +08:00
• e905b4b6cc [Perf][SM70] Run GDN b/a in the QKVZ launch for 9-32 verify rows (#1368) 3 people 2026-11-05 22:45:05 +08:00
• 3255da7bb8 [Perf][SM70] Let K16 producers hand their twin to the next projection (#1367) 3 people 2026-11-05 22:44:44 +08:00
• e098ee11b9 [Bugfix] Composite transparent image inputs before RGB conversion (#1366) 3 people 2026-11-05 21:32:21 +08:00
• d14e383bea [Bugfix][KV Offload] Keep external prefix hits on the Mamba page grid (#1360) LeonandClaude Opus 5.5 2026-11-05 06:30:15 -07:00
• b6be945400 [Bugfix][Frontend] Return output without logprobs when its rows do not match its tokens (#1357) LeonandClaude Opus 5.5 2026-11-05 06:30:06 -07:00
• 0f200e3728 [Bugfix][Loader] Fail fast on incomplete local sharded safetensors checkpoints (#1356) LeonandClaude Opus 5.5 2026-11-05 06:29:57 -07:00
• 9284c41abe [Bugfix] Hash zero multimodal limits into the model compile-cache key (#1355) LeonandClaude Opus 5.5 2026-11-05 06:29:50 -07:00
• 6a4d1f618b [Bugfix][Frontend] Bound the pending window of incremental detokenization (#1354) LeonandClaude Opus 5.5 2026-11-05 06:29:42 -07:00
• 3880f9fcca [Bugfix] Align stop strings with emitted output metadata (#1305) LeonandCodex 2026-11-05 06:29:34 -07:00
• 9433700de3 [Bugfix] Preserve prompt logprobs when merging parallel DELTA outputs (#1318) Leon 2026-11-05 06:29:25 -07:00
• 6e0117f355 [Bugfix] Enforce parallel_tool_calls=false in the required-tool grammar (#1316) LeonandCodex 2026-11-05 06:29:16 -07:00
• fd2062680b [Bugfix] Constrain a named tool without parameters to a JSON object (#1315) LeonandCodex 2026-11-05 06:29:09 -07:00
• 60ec53427c [Bugfix] Honor tool_choice="none" in streaming tool parsing (#1314) LeonandCodex 2026-11-05 06:29:00 -07:00
• fa94145433 [Bugfix] qwen3_coder: stream a call body that completes right after its header (#1313) LeonandCodex 2026-11-05 06:28:52 -07:00
• 5b497c374e [Bugfix] Normalize EXIF orientation when loading images (#1310) LeonandCodex 2026-11-05 06:28:40 -07:00
• b2a9f8b1a3 [Bugfix] Keep non-finite tool arguments JSON-safe (#1309) 3 people 2026-11-05 06:28:33 -07:00
• 87868e0f66 [Bugfix] Preserve large integer tool arguments (#1308) 3 people 2026-11-05 06:28:24 -07:00
• 5ab420dccb [Bugfix] Match loaded runtime filenames before binding cudart (#1307) 3 people 2026-11-05 06:28:17 -07:00
• 1652d0ac42 [Bugfix] Poll shared-memory notifications while readers stay busy (#1306) 3 people 2026-11-05 06:28:08 -07:00
• a9e56e285d [Bugfix] Honor prefix cache skips at external lookup layers (#1304) LeonandOpenAI Codex 2026-11-05 06:28:00 -07:00
• 45251e7aa5 [Bugfix] Restrict the TurboMind AWQ override to AWQ checkpoints (#1303) 3 people 2026-11-05 06:27:53 -07:00
• dc062c170d [Perf][SM70] Qwen3.8 dual-compile lane: hybrid PLE by default when the host can pin it (#1340) yangzhuxinyzxandClaude Opus 5.5 2026-11-05 21:27:35 +08:00
• d2d8e84b73 [Bugfix][Core] Cache output blocks of reasoning requests (#1338) yangzhuxinyzxandClaude Opus 5.5 2026-11-05 21:27:27 +08:00
• 96564ce015 [Misc][SM70] Move full diagnostic dumps to DEBUG (#1364) yangzhuxinyzxandClaude Opus 5.5 2026-11-05 21:27:19 +08:00
• de2568c9a9 [Bugfix][Prefix Cache] Key partial prefix entries on image identity (#1363) yangzhuxinyzxandClaude Opus 5.5 2026-11-05 21:27:12 +08:00
• b2b6b3bb91 docs: refresh WeChat community QR image (#1352) yangzhuxinyzx 2026-11-04 10:44:13 +08:00
• 6db9c816cd [Perf][PLE] Serve FP8 checkpoint PLE rows from the native fast lane (#1343) yangzhuxinyzxandClaude Opus 5.5 2026-11-02 15:11:55 +08:00
• 600aa6c9cd [Bugfix][SM70] Clamp TurboMind split-K to the workspace when dispatch overrides the group count (#1339) yangzhuxinyzxandClaude Opus 5.5 2026-11-02 14:58:57 +08:00
• 4d57473db6 [Frontend][SM70] Report cached prompt tokens in the Qwen3.8 release profile (#1299) yangzhuxinyzxandClaude Opus 5.5 2026-10-29 01:56:36 +08:00
• 6f649847c3 [Perf][SM70] Split-KV for short prefix-prefill chunks over long prefixes (#1298) yangzhuxinyzxandClaude Opus 5.5 2026-10-29 01:56:13 +08:00
• ac06e9771c [Perf][Core][SM70] Partial prefix-cache hits for hybrid Mamba models; agent turns resume at the last message (#1297) yangzhuxinyzxandClaude Opus 5.5 2026-10-29 01:33:08 +08:00
• 5a250ebda7 [Perf][SM70] Qwen3.8-27B profile: 2048-token mamba blocks for prefix-cache hits (#1296) yangzhuxinyzxandClaude Opus 5.5 2026-10-28 22:51:51 +08:00
• 734db62018 [Perf][SM70] Qwen3.8-27B profile: prefill beside resident decoders in full chunks (#1295) yangzhuxinyzxandClaude Opus 5.5 2026-10-28 22:36:26 +08:00
• 6aea01fc6f [Perf][SM70][Distributed] Fused TP4 push all-reduce + Gemma RMSNorm for 16/24/32 verifier rows (#1292) yangzhuxinyzxandClaude Opus 5.5 2026-10-28 21:35:19 +08:00
• 42a8e6fa9a [Perf][SM70][Scheduler] Piecewise graphs for short mixed prefills; keep the mixed budget off the graph regime (#1290) yangzhuxinyzxandClaude Opus 5.5 2026-10-28 21:34:29 +08:00
• e95706cdd7 [Perf][SM70][GDN] BV4 tiles for the two- and three-request DFlash2 verifier update (#1291) yangzhuxinyzxandClaude Opus 5.5 2026-10-28 21:25:30 +08:00
• 9a33cee2ac [Perf][SM70][Spec Decode] Use the DFlash2 draft FP8 projections for 2-4 concurrent requests (#1289) yangzhuxinyzxandClaude Opus 5.5 2026-10-28 19:58:11 +08:00
• 0427e116b9 [Bugfix][Scheduler] Fit the fixed cost of mixed prefill steps instead of charging it to prefill (#1288) yangzhuxinyzxandClaude Opus 5.5 2026-10-28 19:09:49 +08:00
• 3392fec044 [Perf][SM70] 4096-token prefill chunks: native Q4096 79T attention and reserved long-prefix workspaces (#1287) yangzhuxinyzxandClaude Opus 5.5 2026-10-28 18:20:43 +08:00
• abec8ba85e [Bugfix] Keep the pynccl all-reduce out-of-place by default (#1284) yangzhuxinyzxandClaude Opus 5.5 2026-10-28 13:43:01 +08:00
• c59b3c4cbb [Bugfix][Multimodal] Stop the shm processor cache from pinning items for good (#1200) tenotekandClaude Fable 5.1 2026-10-28 10:31:46 +08:00
• 04718eafc9 [Bugfix][QSA] Clamp the ring's cached RoPE positions before they index cos_sin_cache (#1201) tenotekandClaude Fable 5.1 2026-10-28 10:31:25 +08:00
• 089ab495a5 [Bugfix][MLA] Size the chunked-prefill workspace and the indexer prefill buffer by max_num_seqs (#1275) PeuquiandClaude Sonnet 5.5 2026-10-28 03:31:12 +01:00
• 691b6597c9 [Bugfix][Loader] Load index shards in subfolders; read U16 and U32 through direct I/O (#1274) PeuquiandClaude Opus 5.5 2026-10-28 03:31:01 +01:00
• 4d5ee06e2c [Bugfix][SM70] Classify compact top-k/top-p rows on device instead of syncing (#1267) 3 people 2026-10-28 09:30:50 +07:00
• e984902304 [Spec] Backport disable_eagle_block_drop for EAGLE/MTP prefix hits (#1265) jcborroandjcborro 2026-10-27 20:30:34 -06:00
• 4e89c28f86 [Perf] Run the fused GDN speculative metadata launch on Turing too (#1264) PeuquiandClaude Sonnet 5.5 2026-10-28 03:30:20 +01:00
• b6f8980517 [Perf] Do not send a request's token history to the V2 runner every decode step (#1263) PeuquiandClaude Sonnet 5.5 2026-10-28 03:30:08 +01:00
• 29b0590e8c [Bugfix][KV] MLA specs carry the codec of their cache dtype (#1272) PeuquiandClaude Opus 5.5 2026-10-28 03:29:47 +01:00
• e7edbbc905 [Perf][Attention] Pack GQA heads for FA2 multi-token decode verification (#1227) 3 people 2026-10-28 03:29:25 +01:00
• cd74c0e285 [Bugfix][DeepSeek-V4] Restore DSpark decode on SM70 after the per-engine policy migration (#1183) PeuquiandClaude Opus 5.5 2026-10-28 03:29:06 +01:00
• 220953f951 [Perf][SM70][DeepSeek-V4] Keep the cuBLAS decode indexer above 131,072 tokens of context (#1222) PeuquiandClaude Opus 5.5 2026-10-28 03:28:54 +01:00
• 372b7be62f [Bugfix][Distributed] Verify P2P before provisioning compact collectives without custom all-reduce (#1219) PeuquiandClaude Opus 5.5 2026-10-28 03:28:42 +01:00
• 573648d1d2 [Perf][KV Offload] Restore Mamba states where EAGLE GPU hits can reuse them (#1244) Leon 2026-10-27 19:28:28 -07:00
• e0c99ed300 [Perf][FLA] Let the delta_h autotuner lift the ptxas register cap (#1226) 3 people 2026-10-28 03:28:18 +01:00
• b5b7707a12 [Bugfix][KV Offload] Finish reused-block stores before MRv2 zeroes them (#1232) Leon 2026-10-27 19:28:05 -07:00
• fce96323d8 [Bugfix][KV Cache] Zero packed DCP pages from the page base (#1230) Leon 2026-10-27 19:27:37 -07:00
• 40902f3dd3 [Perf][Qwen4Exp] Admit the cuBLAS QSA indexer and exact top-k on Turing (#1229) 3 people 2026-10-28 03:27:17 +01:00
• 0827e0147a [Bugfix][GDN] Keep the traced prefill output projection on the custom-op norm (#1213) PeuquiandClaude Opus 5.5 2026-10-28 03:26:58 +01:00
• cfe5c14806 [Bugfix][Qwen4Exp] Pin the PLE host share at its exact size (#1199) 3 people 2026-10-27 19:26:48 -07:00
• 10157f5970 [Bugfix][Qwen4Exp] Decide the PLE cascade merge when the layer is built (#1234) 3 people 2026-10-28 03:26:37 +01:00
• 7b1cbd1727 [Build][SM70] Fix the Docker source build and keep all SM70 extensions (#1207) LeonandClaude Opus 5.5 2026-10-27 19:26:16 -07:00
• 3bf45b2d86 [Bugfix][Core] Reserve the null KV block during startup admission (#1247) 3 people 2026-10-27 19:26:00 -07:00
• 10b029d845 [Core] Align Mamba replay boundaries without an EAGLE block drop (#1240) PeuquiandClaude Opus 5.5 2026-10-28 03:25:49 +01:00
• 533ebcff03 [Bugfix][Qwen4Exp] Reduce the PLE offload config to DP1 before it is built (#1237) PeuquiandClaude 2026-10-28 03:25:18 +01:00
• 5366819383 [Test] Repair stale fakes in PLE offload and SM70 decode-graph tests (#1214) PeuquiandClaude Opus 5.5 2026-10-28 03:24:54 +01:00
• 9cf2d94a75 [Bugfix][DeepSeek-V4] Keep hc_head finite under float16 for attention-sink rows (#715) 3 people 2026-10-28 03:23:38 +01:00
• 899e9ed331 [Model] DeepSeek-V4.1-Flash on 4x V100: 512K packed KV, faster prefill, API launch (#1277) yangzhuxinyzxandClaude Opus 5.5 2026-10-27 23:33:35 +08:00
• 6a143197e6 [Test] Cover compact hybrid page utilization (#1259) yangzhuxinyzx 2026-10-24 03:12:48 +08:00
• 4e0023d595 [Bugfix] Require a compact hybrid divisor to fill its padded page (#1258) yangzhuxinyzx 2026-10-24 03:03:19 +08:00
• 174e21f387 [Perf] Bound TurboQuant continuation SDPA working memory (#1257) yangzhuxinyzx 2026-10-24 02:06:07 +08:00
• 99ddd4693c [Bugfix] Keep a padded page at least as large as its grown payload (#1256) yangzhuxinyzx 2026-10-24 01:27:01 +08:00
• 54befdc7af [Bugfix] Map NVFP4 in AttentionSpec.layout_cache_dtype (#1255) yangzhuxinyzx 2026-10-24 01:14:55 +08:00
• b32072ac9c [Profile] Add an opt-in SM70 round-cost ledger (#1254) yangzhuxinyzx 2026-10-24 00:45:45 +08:00
• 3bdbb684be [Core] Slice packed GGUF rows before transcoding on the host fallback (#1253) yangzhuxinyzx 2026-10-24 00:36:05 +08:00
• 931eee9fbd [Memory] Bound video-disabled MRoPE caches to the position envelope (#1019) yangzhuxinyzx 2026-10-24 00:12:18 +08:00
• dbc1f4f8fa [Core] Size compact hybrid pages by factor search (#1252) yangzhuxinyzx 2026-10-23 12:06:09 +08:00
• 6b750cd017 [Bugfix] Preserve logical blocks when padding mixed-precision hybrid pages (#1251) yangzhuxinyzx 2026-10-23 12:01:27 +08:00
• 1d7321bf2d Merge pull request #1242 from 1CatAI/claude/flashnext-prefill-merge yangzhuxinyzx 2026-10-22 16:31:53 +08:00
• b4df7b0ea4 [CI] Make the Flash-Next merge pass the repository pre-commit claude 2026-10-22 16:21:39 +08:00
• c420c4e9fd [Style] Apply pre-commit formatting to the concurrency changes claude 2026-10-22 15:29:51 +08:00
• 705aca2120 [Bugfix] Keep GDN replay correct when decodes share a step with prefill claude 2026-10-22 10:53:30 +08:00