Repository navigation

Karmic beta to canonical (Oct 5): rows of #868, through #982 - #894

Closed

voipmonitor wants to merge 496 commits into

dev/karmic-krakenfrom

canonical-candidate/karmic-20260924

Closed

Karmic beta to canonical (Oct 5): rows of #868, through #982#894
voipmonitor wants to merge 496 commits into
dev/karmic-krakenfrom
canonical-candidate/karmic-20260924

This pull request is big! We're only showing the most recent 250 commits

Commits on Sep 23, 2026
- feat(dflash): shard the context projection and draft per vocab shard at TP > 1
  
  MadeBy561
  and
  claude
  committed
  b85ea19
- feat(dflash): support block-FP8 drafters in the fused context projection
  
  MadeBy561
  and
  claude
  committed
  5a98e01
- fix(spec-decode): keep the adaptive verification acceptance estimate recoverable
  
  MadeBy561
  and
  claude
  committed
  addaaa3
- perf(scheduler): let prefill-only steps use max_num_batched_tokens
  
  MadeBy561
  and
  claude
  committed
  d7fadbb
- fix(multimodal): serve MiMo-V2 audio without torchaudio
  
  MadeBy561
  and
  claude
  committed
  6b71316
- feat(mimo): launch MiMo-V2.6-Flash with DFlash and the b12x serving flags
  
  MadeBy561
  and
  claude
  committed
  b04d9bf
- docs(release): add vllm-871 release fragment
  
  MadeBy561
  and
  claude
  committed
  45f0e24
- docs(release): add vllm-872 release fragment
  
  MadeBy561
  and
  claude
  committed
  b7998ab
- docs(release): add vllm-873 release fragment
  
  MadeBy561
  and
  claude
  committed
  3024840
- docs(release): add vllm-874 release fragment
  
  MadeBy561
  and
  claude
  committed
  e8d4aa5
- docs(release): add vllm-875 release fragment
  
  MadeBy561
  and
  claude
  committed
  29e1cf4
- docs(release): add vllm-876 release fragment
  
  MadeBy561
  and
  claude
  committed
  d7b38ca
- docs(release): add vllm-877 release fragment
  
  MadeBy561
  and
  claude
  committed
  0973d8b
- docs(release): add vllm-878 release fragment
  
  MadeBy561
  and
  claude
  committed
  2fdc99d
- docs(release): add vllm-879 release fragment
  
  MadeBy561
  and
  claude
  committed
  18992b1
- docs(release): add vllm-880 release fragment
  
  MadeBy561
  and
  claude
  committed
  bd4cbd7
- docs(release): add vllm-881 release fragment
  
  MadeBy561
  and
  claude
  committed
  64c563c
- chore(release): describe #869
  voipmonitor
  committed
  4a3a8ec
- chore(release): describe #870
  voipmonitor
  committed
  f278b4a
- Merge pull request #869 from local-inference-lab/fix/offloading-hybrid-store-tracking
  voipmonitor
  authored
  a94eea0
- Merge pull request #870 from local-inference-lab/fix/aligned-mamba-external-prefix-hits
  voipmonitor
  authored
  b5d8259
Commits on Sep 24, 2026
- fix(config): let aligned Qwen offload through the atomic-checkpoint gate
  voipmonitor
  committed
  212f46b
- chore(release): describe #883
  voipmonitor
  committed
  3c2dd48
- Merge pull request #883 from local-inference-lab/fix/qwen-aligned-offload-gate
  voipmonitor
  authored
  9e5c2f3
- feat(serve): time requests from HTTP receipt, not only from rendering
  voipmonitor
  committed
  6cf4fc4
- chore(release): identify the receipt timeline change as #884
  voipmonitor
  committed
  08615a6
- Merge pull request #884 from local-inference-lab/diag/request-receipt-timeline
  voipmonitor
  authored
  8a149a1
- [Bugfix][Spec Decode] Stop dummy draft decode steps from writing KV through stale block-table rows (#56734)
  
  ivanium
  authored and
  voipmonitor
  committed
  bca7b17
- fix(qwen): allow full-TP DCP for qwen4_exp checkpoints
  yatesdr
  committed
  5314ef5
- docs(release): identify Qwen full-TP DCP fix as PR 886
  yatesdr
  committed
  d231dda
- chore(release): describe #885
  voipmonitor
  committed
  c3034cd
- test(qwen): check the TP4/DCP4 gate through ModelConfig
  yatesdr
  committed
  4bba478
- Merge pull request #885 from local-inference-lab/fix/dummy-draft-stale-block-tables
  voipmonitor
  authored
  33db77e
- fix: DCP LSE correction for non-power-of-two ranks; pad aligned state rows with NULL_BLOCK_ID
  
  voipmonitor
  and
  claude
  committed
  638f90a
- chore(release): describe #887
  
  voipmonitor
  and
  claude
  committed
  6a35d02
- Pad GLM-5.3 tensor-parallel axes for TP3
  
  voipmonitor
  and
  claude
  committed
  6b30e92
- Run the GLM-5.3 vision tower data-parallel when TP cannot shard it
  
  voipmonitor
  and
  claude
  committed
  ed3e6b3
- Add GLM-5.3 TP3 FP8 decode and low-latency GEMM selection
  
  voipmonitor
  and
  claude
  committed
  920d8c5
- Save every rank's FlashInfer autotune entries and load them consistently
  
  voipmonitor
  and
  claude
  committed
  8cef00e
- chore(release): describe #889
  
  voipmonitor
  and
  claude
  committed
  54be92b
- Merge pull request #887 from local-inference-lab/fix/generic-from-tp3-port
  voipmonitor
  authored
  2f04b10
- Revert #885: dummy draft steps marked every row as padding
  
  voipmonitor
  and
  claude
  committed
  9db8062
- Merge pull request #890 from local-inference-lab/revert/885-dummy-idx-mapping
  voipmonitor
  authored
  de6c7bf
- Merge pull request #889 from local-inference-lab/feat/glm53-tp3
  voipmonitor
  authored
  5a689e3
- chore(release): reference vllm-887 by fragment id in #889
  
  voipmonitor
  and
  claude
  committed
  42e19d0
- Merge pull request #891 from local-inference-lab/fix/fragment-889-requires
  voipmonitor
  authored
  b62b2be
- chore(release): restore the published vllm-885 fragment
  
  voipmonitor
  and
  claude
  committed
  736ef87
- Merge pull request #892 from local-inference-lab/fix/restore-fragment-885
  voipmonitor
  authored
  95720e4
- Sync Karmic beta with dev/karmic-kraken 1794dcf1
  
  voipmonitor
  and
  claude
  committed
  be0ef12
- chore(release): describe the 2026-09-24 canonical sync
  
  voipmonitor
  and
  claude
  committed
  4b57a8a
- Merge pull request #886 from yatesdr/review/qwen4-exp-full-tp-dcp
  voipmonitor
  authored
  4900f3e
- Merge pull request #893 from local-inference-lab/sync/canonical-20260924
  voipmonitor
  authored
  f6496cb
- fix(worker): save startup plans under the pre-profiling fingerprint
  voipmonitor
  committed
  b1e0a40
- Merge pull request #895 from local-inference-lab/fix/startup-plan-fingerprint
  voipmonitor
  authored
  4f66cb9
Commits on Sep 25, 2026
- [Bugfix] Stop workers when an EngineCore that failed to start exits
  yatesdr
  committed
  b39ca56
- docs(release): identify PR 897
  yatesdr
  committed
  27fda44
- feat(glm53): optional online MXFP8 for the vision tower
  
  voipmonitor
  and
  claude
  committed
  e83cbce
- feat(glm53): optional input embedding table in pinned host RAM
  
  voipmonitor
  and
  claude
  committed
  bd1bad8
- feat(distributed): share one PyNCCL communicator across groups over the same ranks
  
  voipmonitor
  and
  claude
  committed
  bfeab71
- test(distributed): cover shared PyNCCL communicators; format
  
  voipmonitor
  and
  claude
  committed
  f4cd9fa
- chore: release fragment vllm-898
  
  voipmonitor
  and
  claude
  committed
  cfd58db
- Merge pull request #898 from local-inference-lab/feat/glm53-kv-memory
  voipmonitor
  authored
  a603af2
- fix(engine): bound the boundary-checkpoint copy wait like execute_model
  
  voipmonitor
  and
  claude
  committed
  d7301dd
- chore: release fragment vllm-903
  
  voipmonitor
  and
  claude
  committed
  2108061
- Merge pull request #903 from local-inference-lab/fix/boundary-copy-rpc-timeout
  voipmonitor
  authored
  d281c90
- feat(core): boundary checkpoints for image and video requests
  
  voipmonitor
  and
  claude
  committed
  72847c5
- chore: release fragment vllm-904
  
  voipmonitor
  and
  claude
  committed
  7c4fea8
- Merge pull request #904 from local-inference-lab/feat/mm-boundary-checkpoints
  voipmonitor
  authored
  9575a0b
Commits on Sep 26, 2026
- [Feature] support bf16 MoE router and mxfp4 MoE for MiMo V2 (#57784)
  
  authored and
  voipmonitor
  committed
  44d23df
- test(mimo): verify router dtype and DFlash value scaling
  
  logprobz
  authored and
  voipmonitor
  committed
  b6b2250
- style(mimo): format router and draft regressions
  
  logprobz
  authored and
  voipmonitor
  committed
  f4928a8
- Compare CPU router logits to an independent FP32 reference
  
  logprobz
  authored and
  voipmonitor
  committed
  c504861
- Merge branch 'pr-831' into test/mimo26-beta
  voipmonitor
  committed
  fef3a3d
- Merge branch 'pr-832' into test/mimo26-beta
  voipmonitor
  committed
  417700c
- Merge branch 'pr-871' into test/mimo26-beta
  voipmonitor
  committed
  a0a5c62
- Merge branch 'pr-872' into test/mimo26-beta
  voipmonitor
  committed
  791ced3
- Merge branch 'pr-873' into test/mimo26-beta
  voipmonitor
  committed
  e95b5bf
- Merge branch 'pr-874' into test/mimo26-beta
  voipmonitor
  committed
  82c9cbb
- Merge branch 'pr-875' into test/mimo26-beta
  voipmonitor
  committed
  5a47e28
- Merge branch 'pr-876' into test/mimo26-beta
  voipmonitor
  committed
  dcd9744
- Merge branch 'pr-877' into test/mimo26-beta
  voipmonitor
  committed
  fe3bae6
- Merge branch 'pr-878' into test/mimo26-beta
  voipmonitor
  committed
  e555862
- Merge branch 'pr-879' into test/mimo26-beta
  voipmonitor
  committed
  548cc4d
- Merge branch 'pr-880' into test/mimo26-beta
  voipmonitor
  committed
  30e4170
- Merge branch 'pr-881' into test/mimo26-beta
  voipmonitor
  committed
  cb995d8
- docs(release): add fragments for vllm-829, vllm-831 and vllm-832
  
  voipmonitor
  and
  claude
  committed
  af94e88
- Merge pull request #905 from local-inference-lab/integration/mimo26-beta
  voipmonitor
  authored
  b31d20a
- Merge pull request #897 from yatesdr/fix/executor-shutdown-on-failed-start
  voipmonitor
  authored
  4b10351
- fix(boundary-checkpoint): full-strength identity for image spans
  
  voipmonitor
  and
  claude
  committed
  f85f095
- docs(release): add vllm-908 release fragment
  
  voipmonitor
  and
  claude
  committed
  b209dcd
- Merge pull request #908 from local-inference-lab/fix/mm-checkpoint-identity
  voipmonitor
  authored
  04c30fa
Commits on Sep 27, 2026
- DS4.1: BF16 sparse attention arithmetic and FP32 logits by default
  
  voipmonitor
  and
  claude
  committed
  40371c7
- docs(release): link the DS4.1 numerics fragment to #911
  
  voipmonitor
  and
  claude
  committed
  534d985
- Merge pull request #911 from local-inference-lab/feat/ds41-bf16-attention-fp32-heads
  voipmonitor
  authored
  31c6c34
- docs(release): DS4.1 numerics fragment uses the fix category
  
  voipmonitor
  and
  claude
  committed
  5d1af9b
- Merge pull request #912 from local-inference-lab/fix/ds41-numerics-fragment-category
  voipmonitor
  authored
  c7b5e86
- DS4.1: share decode page metadata across CUDA graphs
  
  voipmonitor
  and
  claude
  committed
  ce4be0a
- test(ds41): capture decode attention with a fresh forward context
  
  voipmonitor
  and
  claude
  committed
  5da2ca4
- docs(release): link the DS4.1 graph-memory fragment to #918
  
  voipmonitor
  and
  claude
  committed
  5e2fdd6
- Merge pull request #918 from local-inference-lab/fix/ds41-decode-metadata-graph-memory
  voipmonitor
  authored
  1d95593
- test(ds41): refresh V4.1 workspace tests for BF16 attention and L2 prefetch
  
  voipmonitor
  and
  claude
  committed
  3b294c2
- Merge pull request #919 from local-inference-lab/test/ds41-workspace-tests-bf16-l2pf
  voipmonitor
  authored
  4cbfc24
- docs(release): add the missing fragment for test-only #919
  
  voipmonitor
  and
  claude
  committed
  17dae2e
- Merge pull request #920 from local-inference-lab/docs/ds41-workspace-tests-fragment
  voipmonitor
  authored
  84f5605
- [Frontend] Fix the parsing of missing `string=` in DeepSeek V4 (#56271)
  
  authored and
  voipmonitor
  committed
  190268a
- docs(release): attribute the DSML optional string-attribute import
  
  voipmonitor
  and
  claude
  committed
  574e0fe
- docs(release): link the DSML import fragment to #921
  
  voipmonitor
  and
  claude
  committed
  76554c1
- Merge pull request #921 from local-inference-lab/port/upstream-56271-dsml-optional-string
  voipmonitor
  authored
  eba11af
Commits on Sep 28, 2026
- fix(worker): map circular-buffer groups into their ring block
  
  MadeBy561
  and
  claude
  committed
  dfeb396
- docs(release): record DeepSeek-V4.1 compressor ring fix
  
  MadeBy561
  and
  claude
  committed
  4d00743
- fix(worker): map circular-buffer groups into their ring block
  
  authored and
  voipmonitor
  committed
  7064287
- docs(release): record DeepSeek-V4.1 compressor ring fix
  
  authored and
  voipmonitor
  committed
  8be242a
- Merge pull request #927 from local-inference-lab/port/pr-926-compressor-ring
  voipmonitor
  authored
  eeb359d
- fix(kv-cache): give circular-buffer groups a worker slot mapping
  
  MadeBy561
  and
  claude
  committed
  5eec8ea
- docs(release): record DeepSeek-V4.1 compressor ring writes
  
  MadeBy561
  and
  claude
  committed
  7b935cb
- fix(executor): gather every rank's reply without waiting in rank order
  
  tobymao
  authored and
  voipmonitor
  committed
  b98ece7
- fix(executor): read a gathered reply only once it has arrived
  
  tobymao
  authored and
  voipmonitor
  committed
  34d0206
- Preserve external checkpoint restores through capacity-constrained admission
  
  ktsaou
  authored and
  voipmonitor
  committed
  4c5e427
- Link checkpoint admission release note to PR 922
  
  ktsaou
  authored and
  voipmonitor
  committed
  ae02fd7
- fix(executor): surface a failed rank during single-rank RPCs
  
  voipmonitor
  and
  claude
  committed
  64f0d17
- fix(serve): exit with 1 when the server stops because its engine failed
  
  voipmonitor
  and
  claude
  committed
  a8ae189
- docs(release): record the fast exit on a failed rank
  
  voipmonitor
  and
  claude
  committed
  eb9fec3
- fix(serve): count only a failed engine exit as an engine failure
  
  voipmonitor
  and
  claude
  committed
  1d5a74c
- fix(kv-cache): hold import reservations only against new admissions
  
  voipmonitor
  and
  claude
  committed
  3f94f17
- fix(scheduler): keep queue order and step cost while imports are pending
  
  voipmonitor
  and
  claude
  committed
  0dd394d
- fix(scheduler): release restored imports on a forced prefix-cache reset
  
  voipmonitor
  and
  claude
  committed
  7023056
- docs(release): describe fair, bounded checkpoint-restore admission
  
  voipmonitor
  and
  claude
  committed
  f9082f2
- Merge pull request #928 from local-inference-lab/fix/worker-failure-fast-exit
  voipmonitor
  authored
  8586695
- docs(release): link the restore-admission fragment to PR 929
  
  voipmonitor
  and
  claude
  committed
  1ac6c2d
- Merge pull request #929 from local-inference-lab/fix/restore-admission-followup
  voipmonitor
  authored
  6ea2fcf
- fix(structured-output): keep MTP draft rows constrained after a skipped step
  
  voipmonitor
  and
  claude
  committed
  0928737
- fix(cudagraph): measure sampled graphs after the warmup cache is flushed
  
  voipmonitor
  and
  claude
  committed
  1c4a925
- docs(release): describe the CUDA graph memory estimate fix
  
  voipmonitor
  and
  claude
  committed
  e7fcc19
- docs(release): describe the structured-output draft hand-off fix
  
  voipmonitor
  and
  claude
  committed
  d58d377
- fix(parser): return GLM reasoning and content with generated whitespace
  
  voipmonitor
  and
  claude
  committed
  eec2382
- docs(release): record GLM parser whitespace preservation
  
  voipmonitor
  and
  claude
  committed
  b5dedaf
- test(worker): stub the draft hand-off in the logits-only runner test
  
  voipmonitor
  and
  claude
  committed
  c9181eb
- revert: the DeepSeek-V4.1 compressor-ring change (#927) had no effect
  
  voipmonitor
  and
  claude
  committed
  3fbef72
- docs(release): correct the DeepSeek-V4.1 drift note
  
  voipmonitor
  and
  claude
  committed
  0aa6de7
- Merge pull request #933 from local-inference-lab/revert/927-ring-no-effect
  voipmonitor
  authored
  1d75cf2
- fix(tokenizers): max_token_id is vocab_size - 1, not vocab_size
  
  voipmonitor
  and
  claude
  committed
  2962205
- docs(release): record the max_token_id fix
  
  voipmonitor
  and
  claude
  committed
  f9cf8e2
- Merge pull request #934 from local-inference-lab/fix/max-token-id-off-by-one
  voipmonitor
  authored
  40a89ca
- test(parser): cover GLM exact-history turns through the serving API
  
  voipmonitor
  and
  claude
  committed
  7b2fc4e
- Merge pull request #931 from local-inference-lab/fix/726-mtp-grammar
  voipmonitor
  authored
  635160c
- Merge pull request #932 from local-inference-lab/fix/response-checkpoint-template-reuse
  voipmonitor
  authored
  590be52
Commits on Sep 29, 2026
Commits on Sep 30, 2026
- fix(executor): keep replies that answer a later RPC
  
  voipmonitor
  and
  claude
  committed
  9ce5adb
- Merge pull request #950 from local-inference-lab/fix/rpc-reply-held-for-later-rpc
  voipmonitor
  authored
  5df66ad
- chore(release): wheels require CUTLASS DSL 4.7.1 with B12X
  
  voipmonitor
  and
  claude
  committed
  16c5236
- Merge pull request #951 from local-inference-lab/chore/cutlass-dsl-4.7.1
  voipmonitor
  authored
  02ab538
- Release dense MLA priming operands after synchronous preparation
  voipmonitor
  committed
  371ca44
- Load Kimi-K3 X4T experts through the lossless MXFP4 reader
  voipmonitor
  committed
  c390adb
- Document transient MLA preparation ownership for runtime releases
  
  voipmonitor
  and
  codex
  committed
  5f8a575
- Declare Kimi X4T loader compatibility for runtime releases
  
  voipmonitor
  and
  codex
  committed
  120a909
- fix(release): install CUTLASS DSL 4.7.1 in the wheel build environment
  
  voipmonitor
  and
  claude
  committed
  da9fc45
- Merge pull request #954 from local-inference-lab/fix/wheel-builder-cutlass-4.7.1
  voipmonitor
  authored
  a2b44f5
Commits on Oct 1, 2026
- Load losslessly compressed GLM NVFP4 scales with native A4 experts
  
  Martin Vit
  and
  codex
  committed
  e530531
- Expose MXFP4-CSF and NVFP4-CSF with compatible checkpoint loading
  
  Martin Vit
  and
  codex
  committed
  77e3dd8
- Load canonical CSF models with retained native quantization
  voipmonitor
  committed
  672c8f6
- Document canonical CSF loading for GLM and Qwen TP2
  voipmonitor
  committed
  096c703
- feat(moe): integrate native B12X MXFP8 MTP experts
  
  authored and
  voipmonitor
  committed
  c65181b
- test(moe): disable established MXFP8 backends in b12x fallback test
  
  randomvariable
  authored and
  voipmonitor
  committed
  714e851
- feat(moe): prepare B12X MXFP8 experts at 32-aligned intermediate sizes
  
  randomvariable
  authored and
  voipmonitor
  committed
  858af5c
- docs(b12x): state the MXFP8 per-rank intermediate requirement as % 32
  
  randomvariable
  authored and
  voipmonitor
  committed
  71dbc71
- docs(moe): document B12X MXFP8 adapter and test helpers
  
  randomvariable
  authored and
  voipmonitor
  committed
  66ab251
- test(moe): pin B12X MXFP8 selection gates and the prepared FC1 order
  
  voipmonitor
  and
  claude
  committed
  fb5cacb
- docs(release): describe B12X MXFP8 MTP draft experts
  
  voipmonitor
  and
  claude
  committed
  69ba12d
- Merge pull request #958 from local-inference-lab/port/beta-b12x-mxfp8-mtp-experts
  voipmonitor
  authored
  8531413
- fix(scheduler): give prefill lanes to queued requests only while admissible
  
  voipmonitor
  and
  claude
  committed
  6f4b644
- docs(release): describe the prefill-lane admission stall fix
  
  voipmonitor
  and
  claude
  committed
  01ad5ad
- Merge pull request #959 from local-inference-lab/fix/prefill-lanes-need-admission
  voipmonitor
  authored
  980d84e
- feat(qwen4exp): share the mapped-host PLE table between same-host processes
  
  committed
  348f4fa
- Merge vllm#841
  voipmonitor
  committed
  585b65f
- Merge vllm#858
  
  voipmonitor
  and
  claude
  committed
  d57a8bb
- Merge vllm#937
  voipmonitor
  committed
  65ca3ed
- Merge vllm#949
  voipmonitor
  committed
  e4dc612
- Merge vllm#956
  voipmonitor
  committed
  2c25fa0
- Merge vllm#953
  voipmonitor
  committed
  327d9cc
- Merge vllm#846
  
  voipmonitor
  and
  claude
  committed
  5d5cec9
- docs(release): describe the shared PLE table
  
  voipmonitor
  and
  claude
  committed
  5daabd5
- Merge pull request #961 from local-inference-lab/feat/qwen-shared-ple-table
  voipmonitor
  authored
  4a379ed
- Merge integration/karmic-kraken-beta (shared PLE table, #961)
  voipmonitor
  committed
  9886123
- Own CSF checkpoint discovery in vLLM and supply generic tensor sources
  
  voipmonitor
  and
  codex
  committed
  4d11fdc
- Own CSF checkpoint slicing and use B12X weight preparation
  voipmonitor
  committed
  80c4201
- Match DS4.1 CSF expert activations to native MXFP8 policy
  voipmonitor
  committed
  a5430f5
Commits on Oct 2, 2026
- Merge CSF loader ownership and native expert precision into beta composition
  voipmonitor
  committed
  c7acb24
Commits on Oct 3, 2026
- Merge pull request #963 from local-inference-lab/port/beta-906-csf
  
  voipmonitor
  and
  claude
  committed
  1ae4c63
- Quantize exact GLM Spark projections and MTP experts on load
  
  voipmonitor
  and
  claude
  committed
  09da98a
- Link GLM online weight release metadata to PR 965
  voipmonitor
  committed
  ae16804
- Merge pull request #965 from local-inference-lab/feat/glm-spark-online-weights
  
  voipmonitor
  and
  claude
  committed
  1286ae9
Commits on Oct 4, 2026
- b12x MoE: opt-in NVFP4-activation prefill for W4A16 experts, decode rows stay W4A16
  
  voipmonitor
  and
  claude
  committed
  64303cb
- tests: host contract of the opt-in b12x W4A16 A4 prefill (threshold, decode rows, split)
  
  voipmonitor
  and
  claude
  committed
  9e60da6
- release: the vLLM A4 prefill change requires the b12x feature fragment
  
  voipmonitor
  and
  claude
  committed
  1b6c2a3
- release: link the A4 prefill fragment to #970
  
  voipmonitor
  and
  claude
  committed
  d2e74c0
- b12x MoE: lint fixes for the A4 prefill split
  
  voipmonitor
  and
  claude
  committed
  8fa22e4
- release: A4 prefill fragment with the measured GLM numbers
  
  voipmonitor
  and
  claude
  committed
  9e4d0f6
- Merge pull request #970 from local-inference-lab/feat/b12x-w4a16-a4-prefill
  voipmonitor
  authored
  3c6c8e4
- Serve lossless MXFP4-CSF DeepSeek-V4-Flash and Vision-Exp checkpoints
  
  voipmonitor
  and
  claude
  committed
  a628c6a
- release: link the DS4-Flash MXFP4-CSF fragment to #971
  
  voipmonitor
  and
  claude
  committed
  2338b9c
- Merge pull request #971 from local-inference-lab/feat/mxfp4-csf-ds4-flash
  voipmonitor
  authored
  fba0d19
- qwen4_exp: resolve the PLE storage dtype from an FP4-CSF checkpoint root
  
  voipmonitor
  and
  claude
  committed
  44100d3
- release: link the PLE dtype fragment to #974
  
  voipmonitor
  and
  claude
  committed
  87b54c5
- Merge pull request #974 from local-inference-lab/fix/qwen-ple-dtype-csf
  voipmonitor
  authored
  7f17287
- modelopt: zero-initialize NVFP4 MoE input scales
  
  voipmonitor
  and
  claude
  committed
  d7af3ad
- release: fragment for the zero-initialized NVFP4 MoE input scales (#976)
  
  voipmonitor
  and
  claude
  committed
  dbece92
- Merge pull request #976 from local-inference-lab/fix/nvfp4-moe-unloaded-input-scales
  voipmonitor
  authored
  be03c27
- b12x MoE: choose the A4 prefill by row type, not by call size
  
  voipmonitor
  and
  claude
  committed
  df3d983
- release: fragment for the row-type A4 prefill (#977)
  
  voipmonitor
  and
  claude
  committed
  109d044
- Merge pull request #977 from local-inference-lab/feat/b12x-a4-prefill-row-type
  voipmonitor
  authored
  ad3743e
- NVFP4-CSF: expand the next layer's scales during its attention
  
  voipmonitor
  and
  claude
  committed
  53c7474
- release: fragment for the NVFP4-CSF scale prefetch (#978)
  
  voipmonitor
  and
  claude
  committed
  fcad04e
- Merge pull request #978 from local-inference-lab/perf/b12x-csf-scale-prefetch
  voipmonitor
  authored
  ff292a9
- [Worker] Load weights without the split limit under expandable segments
  
  voipmonitor
  and
  claude
  committed
  f2d1272
- release: fragment for loading without the split limit (#980)
  
  voipmonitor
  and
  claude
  committed
  ffb0d84
- Merge pull request #980 from local-inference-lab/fix/load-split-expandable-segments
  voipmonitor
  authored
  00a33e2
Commits on Oct 5, 2026
- MXFP4-CSF: load DeepSeek-V4.1-Flash checkpoints at TP3
  Defilan
  committed
  65c16dd
- lil: link the release-note fragment to #982
  Defilan
  committed
  3b488b1
- [b12x] Clarify A4 prefill logs for FP4-CSF and draft MoE layers
  
  voipmonitor
  and
  claude
  committed
  7f37b59
- Merge pull request #983 from local-inference-lab/fix/b12x-a4-prefill-log
  voipmonitor
  authored
  8a5d364
- Merge pull request #982 from Defilan/feat/csf-tp3
  voipmonitor
  authored
  f1c2508

Commits

Commits on Sep 23, 2026

Commits on Sep 24, 2026

Commits on Sep 25, 2026

Commits on Sep 26, 2026

Commits on Sep 27, 2026

Commits on Sep 28, 2026