Repository navigation

server: GLM tool calls stream their arguments, and one-call streams match the whole reply (upstream #1975, #1396) - #2

Merged

devopsfortunato merged 5723 commits into

mainfrom

card/t_14cd2239

Nov 3, 2026

Merged

server: GLM tool calls stream their arguments, and one-call streams match the whole reply (upstream #1975, #1396)#2
devopsfortunato merged 5723 commits into
mainfrom
card/t_14cd2239

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

Commits on Nov 2, 2026
- exl3 cuda: review fixes - plan()'s launches, group reserves after the MLP, group counters through grownCounters, the replacements in the kernel contract
  
  BHCC2025
  authored and
  ashhart
  committed
  60db57e
- exl3 cuda: every mid-M, linear_wc and group form the dispatch reaches in the kernel contract
  
  BHCC2025
  authored and
  ashhart
  committed
  9c05804
- exl3 cuda: the mid-M and linear_wc tables in exl3_midm.zig, so exl3.zig stays under 600 lines on main
  
  BHCC2025
  authored and
  ashhart
  committed
  a6a1c64
- kernel contract: the EXL3 set joins the family sets, so contract_tests.zig stays under 600 lines
  
  BHCC2025
  authored and
  ashhart
  committed
  dfc4f21
- docs: the EXL3 recipe lists TF_EXL3_PDL=0, plain launches for the rotations and the linear
  
  BHCC2025
  authored and
  ashhart
  committed
  b24a17b
- kernel contract: the Flash Next and EXL3 sets side by side after the rebase
  ashhart
  committed
  1e1145c
- changelog: #1467 under Unreleased
  ashhart
  committed
  21c8a0f
- qwen27 DFlash: compile the drafter head's lane pipeline at load
  
  kieranscotney
  authored and
  ashhart
  committed
  766c505
- Test the rebased DFlash head cache against prompt and decode launches
  
  ashhart
  and
  kieranscotney
  committed
  cd2e054
- changelog: #1734 under Unreleased
  ashhart
  committed
  0693d4a
- lane host: opt-in shortest remaining work first between a long prompt and running replies
  
  gilby
  authored and
  ashhart
  committed
  554dd6a
- lane host srpt: honour --decode-share (its share when the prompt goes first, at most the floor otherwise; 0 stays whole)
  
  gilby
  authored and
  ashhart
  committed
  e660881
- changelog: #1876 under Unreleased
  ashhart
  committed
  dae3743
- starlight: one governed endpoint for TensorFold fleets
  ashhart
  committed
  cb2b292
- changelog: #1922 under Unreleased
  ashhart
  committed
  6612d18
- starlight: mdns reads browse answers only after the query returns
  ashhart
  committed
  3392389
- starlight: admission makes its fallback waker once
  ashhart
  committed
  1bc80d2
- gemma4: the XPU runner's shard header is read with core's headerSize and parseHeader
  
  jschmied
  authored and
  ashhart
  committed
  89314c6
- build: test-draft-xpu-host and zig build test run gemma4/xpu_dev.zig's tests
  
  jschmied
  authored and
  ashhart
  committed
  2aefc37
- changelog: #1905 under Unreleased
  ashhart
  committed
  d114693
- fix(tools): fail incomplete image benchmark sessions
  
  garvrao80
  authored and
  ashhart
  committed
  88b0d4b
- changelog: #1916 under Unreleased
  ashhart
  committed
  805f737
- flashnext: cuda_embed reads its header with core's headerSize and parseHeader, so bytes match each shape
  
  jschmied
  authored and
  ashhart
  committed
  5ee5bb7
- changelog: #1917 under Unreleased
  ashhart
  committed
  45fcab6
- flashnext cuda: a shared round's attention in one launch, behind TF_FLASHNEXT_ATTN_LANES (#1739)
  
  optimisme
  authored and
  ashhart
  committed
  9d1ffad
- changelog: #1739's line carries its number and thanks
  ashhart
  committed
  c0842b5
- zig fmt (Zig 0.17 spellings) on #1739's files
  ashhart
  committed
  9bb51cd
- qwen4 host test: run #1739's merge body (main's zig CI)
  ashhart
  committed
  72784e2
- core: direct_gather reads a loader's mapped views past the page cache; nvfp4_slice.shapes checks without reading bytes
  
  jschmied
  authored and
  ashhart
  committed
  43e23ac
- flashnext cuda: experts load past the page cache, and a load keeps its shards open instead of reparsing them (#1873)
  
  jschmied
  authored and
  ashhart
  committed
  133c789
- flashnext cuda: release() keeps the n-gram shards' pages, a layer past one gather's shards reads through the mapping, packed chunks' views cleared
  
  jschmied
  authored and
  ashhart
  committed
  0458172
- changelog: #1901 under Unreleased
  ashhart
  committed
  2f339ab
- nemotron: kv_pages grow and keep inline; debug builds assert each pass's rows sit in the lane's pages (#1293)
  
  jschmied
  authored and
  ashhart
  committed
  4675e47
- changelog: #1914 under Unreleased
  ashhart
  committed
  b963f32
- core: add a bounded GGUF v3 reader
  ashhart
  committed
  469c14c
- core: repack finite GGUF blocks into compact affine weights
  ashhart
  committed
  72bf8ed
- changelog: core: a GGUF reader and an exact repack of GGUF's quantized blocks (#1920, steps 1-2)
  ashhart
  committed
  6912945
- drama3 cuda: reuse a voice reference's prefill across requests
  
  jayleaton
  authored and
  ashhart
  committed
  56c8f58
- tools: drama_reference_cache_check.py, the GPU check for the Drama reference cache
  
  jayleaton
  authored and
  ashhart
  committed
  0778b4d
- drama3 cuda: the reference cache is opt-in, its copied K and V rows tested against the kernel's layout
  
  jayleaton
  authored and
  ashhart
  committed
  72f4deb
- server: the Living Weights view reads a shard's header with core, its sums and rows checked
  
  jschmied
  authored and
  ashhart
  committed
  0a83dd5
- changelog: #1923 under Unreleased
  ashhart
  committed
  56f9113
- fix(qwen35/hip): time shared rounds at the served streams and their row buckets
  
  AbrahamCete
  authored and
  ashhart
  committed
  1a46d6f
- fix(qwen35/hip): time forwards of eight rows or more once
  
  AbrahamCete
  authored and
  ashhart
  committed
  b76151e
- fix(qwen35/hip): time each shared bucket at its own rows up to batch_rows
  
  AbrahamCete
  authored and
  ashhart
  committed
  cc14ed2
- changelog: #1830 under Unreleased
  ashhart
  committed
  3609c1e
- refactor(cache): move the prompt cache Rules into their own file
  
  laudney
  authored and
  ashhart
  committed
  b31a399
- feat(cache): let a family count keyed prompts as repeats
  
  laudney
  authored and
  ashhart
  committed
  a85ad8f
- changelog: #1929 under Unreleased
  ashhart
  committed
  94cce45
- feat(hip): add rocBLAS BF16 projections with FP32 sums
  
  laudney
  authored and
  ashhart
  committed
  ac394f7
- changelog: #1930 under Unreleased
  ashhart
  committed
  4de3944
- docs: Flash Next on a Strix Halo iGPU, three checkpoints with speed and quality
  
  gilby
  authored and
  ashhart
  committed
  1cdc24d
- docs: the recipe says oQ4e is the quality pick and what the uniform 4-bit checkpoint costs on tasks
  
  gilby
  authored and
  ashhart
  committed
  3fb7fe4
- changelog: #1747 under Unreleased
  ashhart
  committed
  7193214
- Enable GitHub Sponsors for TensorFold
  ashhart
  committed
  02441c3
- fix(tokenizer): support checkpoint compatibility normalization
  
  laudney
  authored and
  ashhart
  committed
  e014aa7
- fix(tokenizer): decode normalized added tokens by their spelling
  
  laudney
  authored and
  ashhart
  committed
  31dec64
- changelog: #1931 under Unreleased
  ashhart
  committed
  81edd82
- feat(server): let a family declare replies without reasoning
  
  laudney
  authored and
  ashhart
  committed
  f7484b4
- test(server): check the reasoning hook without pinning the registry
  
  laudney
  authored and
  ashhart
  committed
  cd75cd3
- changelog: #1933 under Unreleased
  ashhart
  committed
  933b969
- compact: read a memory marker from its text parts
  
  ppardus
  authored and
  ashhart
  committed
  c3fbea7
- compact: start a memory note after a carriage return
  
  ppardus
  authored and
  ashhart
  committed
  15fa909
- compact: drop a carriage return from a remembered file path
  
  ppardus
  authored and
  ashhart
  committed
  3784ff9
- compact: leave a generated section out after carriage returns
  
  ppardus
  authored and
  ashhart
  committed
  9cd3763
- cortheon: count the space before a fallback word toward its cap
  
  ppardus
  authored and
  ashhart
  committed
  9dc6694
- cortheon: read a shell command that is not the first argument
  
  ppardus
  authored and
  ashhart
  committed
  2fa40c3
- cortheon: drop a partial character at the end of a text cap
  
  ppardus
  authored and
  ashhart
  committed
  2845868
- changelog: #1939 under Unreleased
  ashhart
  committed
  648c4ce
- feat: add stable profile registry and scoped startup reporting
  
  garvrao80
  authored and
  ashhart
  committed
  246d043
- changelog: #1944 under Unreleased
  ashhart
  committed
  16d95cc
- fix(qwen cuda): validate DFlash2 sliding attention declarations
  
  garvrao80
  authored and
  ashhart
  committed
  1173a00
- test(qwen cuda): declare sliding attention in cost cache fixtures
  
  garvrao80
  authored and
  ashhart
  committed
  46aa626
- changelog: #1946 under Unreleased
  ashhart
  committed
  8c3bb8a
- hip: 4096-row prompt chunks by default on the RDNA 3.5 iGPUs
  
  gilby
  authored and
  ashhart
  committed
  aec2bad
- hip: keep 1024-row prompt chunks under --tp unless asked
  
  gilby
  authored and
  ashhart
  committed
  c6cd42f
- changelog: #1651 under Unreleased
  ashhart
  committed
  bd7b066
- perf(hip): narrowest stream columns at RDNA3.5's one- and two-row launches
  
  AbrahamCete
  authored and
  ashhart
  committed
  a3f7889
- refactor(hip): the stream tile's row tile and block width from stream_pick
  
  AbrahamCete
  authored and
  ashhart
  committed
  2cc2ef9
- refactor(hip): assert at comptime that every stream tile width divides 8
  
  AbrahamCete
  authored and
  ashhart
  committed
  cb409d9
- changelog: #1855 under Unreleased
  ashhart
  committed
  089a18d
- ci: run the app and mobile checks only when app/, README.md or their workflow change
  ashhart
  committed
  f32b725
- fix(lanes): release a stream that finishes on its first token
  
  laudney
  authored and
  ashhart
  committed
  ed39acc
- feat(core): add the KV admission ledger
  
  laudney
  authored and
  ashhart
  committed
  3eb4352
- feat(core): admit lane jobs against a byte ledger
  
  laudney
  authored and
  ashhart
  committed
  1721173
- refactor(cuda): decide the stream count in cuda_streams.zig
  
  laudney
  authored and
  ashhart
  committed
  c731947
- feat(cuda): serve per-request families from the KV ledger
  
  laudney
  authored and
  ashhart
  committed
  f8656d1
- docs(core): state the per-request contract on its fields
  
  laudney
  authored and
  ashhart
  committed
  da27d97
- fix(cuda): size the ledger's prompt cache after every stream
  
  laudney
  authored and
  ashhart
  committed
  5bab525
- changelog: #1945 under Unreleased
  ashhart
  committed
  33429c3
- feat(build): check prebuilt images for their one SASS profile
  
  laudney
  authored and
  ashhart
  committed
  f3bce2a
- feat(build): build one stripped server per GPU with -Dgpu
  
  laudney
  authored and
  ashhart
  committed
  ae40e5a
- docs: document the per-GPU product build
  
  laudney
  authored and
  ashhart
  committed
  0e13f76
- fix(cuda): refuse a GPU other than the one -Dgpu built for
  
  laudney
  authored and
  ashhart
  committed
  3dc9693
- docs(runbook): place the one-GPU build after the archive text
  
  laudney
  authored and
  ashhart
  committed
  4193857
- fix(build): defer -Dgpu refusals to the install step
  
  laudney
  authored and
  ashhart
  committed
  bba4c38
- fix(cuda): name a -Dgpu in a refusal only for an SM it builds
  
  laudney
  authored and
  ashhart
  committed
  405abe1
- changelog: #1936 under Unreleased
  ashhart
  committed
  5ef8242
- request_log: the others-can-read test sets its file's mode after creating it, so it holds under any umask
  
  jishnuvenugopal
  authored and
  ashhart
  committed
  8b8deb8
- changelog: #1900 under Unreleased
  ashhart
  committed
  b7a7230
- cortheon: own words find each closer in one pass, so unclosed tags no longer rescan the message
  
  jishnuvenugopal
  authored and
  ashhart
  committed
  2569e9b
- cortheon fuzz: own words match the rescan they replace on every input, and read long ones in linear time
  
  jishnuvenugopal
  authored and
  ashhart
  committed
  b9e4a8f
- changelog: #1898 under Unreleased
  ashhart
  committed
  6c74670
- cortheon: read a GitHub file when its host's letters differ
  
  ppardus
  authored and
  ashhart
  committed
  16bf734
- cortheon: read a GitHub file from its www host
  
  ppardus
  authored and
  ashhart
  committed
  ba3750b
- cortheon: notice a URL written with a capital scheme
  
  ppardus
  authored and
  ashhart
  committed
  3c26c7a
- cortheon: keep a link broken by a carriage return
  
  ppardus
  authored and
  ashhart
  committed
  d0c7848
- cortheon: keep a search result whose scheme is capitalised
  
  ppardus
  authored and
  ashhart
  committed
  15b688d
- changelog: #1949 under Unreleased
  ashhart
  committed
  8e35d52
- cortheon: the unclosed-tags linear-time test reads linear_time.Fastest's fields
  ashhart
  committed
  5fe74eb
Commits on Nov 3, 2026
- fix(cache): retain exact image identities in prompt states
  
  laudney
  authored and
  ashhart
  committed
  1db545b
- changelog: #1962 under Unreleased
  ashhart
  committed
  2c82d4f
- fix(hip): an integrated GPU's budget counts its firmware VRAM carve-out
  
  llstrk
  authored and
  ashhart
  committed
  0e7c216
- fix(hip): an integrated GPU counts its own memory only when sysfs says it is dedicated VRAM
  
  llstrk
  authored and
  ashhart
  committed
  c31b032
- fix(hip): a carve-out's total must match the dedicated VRAM, and the reserve text names the discrete floor
  
  llstrk
  authored and
  ashhart
  committed
  e16790e
- changelog: #1963 under Unreleased
  ashhart
  committed
  3437600
- fix(tools): require observed output for cache benchmark qualification
  
  garvrao80
  authored and
  ashhart
  committed
  318d1a5
- changelog: #1967 under Unreleased
  ashhart
  committed
  8d8e2d7
- docs: Qwen3.8-27B on a 48 GB M5 Pro Mac mini with 1.2.1, drafted and serial
  
  pauldodd123
  authored and
  ashhart
  committed
  16d4466
- changelog: #1947 under Unreleased
  ashhart
  committed
  8914289
- moss_image: the schedule from scheduler_config.json natively, bit for bit MOSSImageScheduler's
  
  BobClawblaw
  authored and
  ashhart
  committed
  a218af5
- changelog: #1670 under Unreleased
  ashhart
  committed
  d8aeab9
- qwen3_5 cuda: deferred keeps opt-in (TF_QWEN_KEEP_LATE=1) until their copy's INVALID_VALUE is fixed
  ashhart
  committed
  755db15
- xpu: a server without DPAS leaves the Qwen3.8-27B runner and Gemma 4 E4B out, and nem_prow gets a 4 x 4 tile image for Gen12LP
  
  jschmied
  authored and
  ashhart
  committed
  0e82346
- xpu: the Gen12LP prompt image takes 8 x 4 tiles (nem_prow_gen12)
  
  jschmied
  authored and
  ashhart
  committed
  10453b0
- changelog: #1692 under Unreleased
  ashhart
  committed
  94ab30f
- feat(qwen3_5): map an EXL3 checkpoint's embedding from host memory only when the card has no room
  
  diwakergupta
  authored and
  ashhart
  committed
  c7ee94f
- changelog: #1678 under Unreleased
  ashhart
  committed
  e854c28
- Issue forms and a sharper pull request template
  
  jayleaton
  authored and
  ashhart
  committed
  41d6c89
- Template and proposal wording from review
  
  jayleaton
  authored and
  ashhart
  committed
  b15a95b
- changelog: #1596 under Unreleased
  ashhart
  committed
  21183f7
- flashnext cuda: prompt passes with the same bits and less work, and a prompt scratch sized by use
  
  Arminova
  authored and
  ashhart
  committed
  2a92fb6
- changelog: #1764 under Unreleased
  ashhart
  committed
  e44c670
- flashnext cuda: cuda_batch.zig within 600 lines on current main (#1764's lean check)
  ashhart
  committed
  9693706
- hip: load a shared expert wider than the routed experts beside the stack
  
  gilby
  authored and
  ashhart
  committed
  06eca5d
- zig fmt
  
  gilby
  authored and
  ashhart
  committed
  8530e2b
- hip: register tf_moe_combine_gate with the kernel contract
  
  gilby
  authored and
  ashhart
  committed
  b51ab36
- qwen3_5: refuse a shared expert the HIP index leaves beside the routed stack
  
  gilby
  authored and
  ashhart
  committed
  2ae3ae4
- hip: test tf_moe_combine_gate against tf_moe_combine byte for byte
  
  gilby
  authored and
  ashhart
  committed
  efd18a9
- changelog: #1612 under Unreleased
  ashhart
  committed
  dacd40a
- ling: grouped expert routing on Metal, with a float64 reference
  
  ThinkOffApp
  authored and
  ashhart
  committed
  4f2c5e7
- ling: contract import and Metal limits launcher for tf_ling_route
  
  ThinkOffApp
  authored and
  ashhart
  committed
  70d1739
- ling: host tests that pin sigmoid, group score, keep-4 and f64 weights
  
  ThinkOffApp
  authored and
  ashhart
  committed
  41db4e2
- ling: reject padded picks and tighten the Metal weight bound to 1e-6
  
  ThinkOffApp
  authored and
  ashhart
  committed
  368941d
- ling: GPU cases for tied groups and a planted NaN on flash
  
  ThinkOffApp
  authored and
  ashhart
  committed
  710f936
- ling: unique picks after NaN, and tests for the last mutation gaps
  
  ThinkOffApp
  authored and
  ashhart
  committed
  19d1357
- ling: distinct-ids on the float64 reference, and the NaN weight rule
  
  ThinkOffApp
  authored and
  ashhart
  committed
  5635e5d
- ling: empty picks are id 0 with weight 0; Shape.of, Shape.check and the bias refusal
  
  ThinkOffApp
  authored and
  ashhart
  committed
  7539909
- ling: route.zig doc comments one line each, within 120 columns
  
  ThinkOffApp
  authored and
  ashhart
  committed
  c2cd15a
- changelog: #1836 under Unreleased
  ashhart
  committed
  bfc374f
- test(hip): pin the planned decode score's bits to its lane order
  
  AbrahamCete
  authored and
  ashhart
  committed
  54fb92e
- perf(hip): the planned decode score takes four keys a warp on long spans, two on short ones
  
  AbrahamCete
  authored and
  ashhart
  committed
  2b4572d
- changelog: #1932 under Unreleased
  ashhart
  committed
  3d22922
- glm cuda: TF_GLM_KV_HOST keeps the KV pages' latent rows in pinned host memory
  
  russellmm
  authored and
  ashhart
  committed
  4bddc94
- glm cuda: the last rank to map TF_GLM_KV_HOST_SHARED's page file unlinks it
  
  russellmm
  authored and
  ashhart
  committed
  160063d
- glm cuda: TF_GLM_KV_HOST_SHARED's /dev/shm pages build on Linux only
  
  russellmm
  authored and
  ashhart
  committed
  ce42bcd
- glm cuda: TF_GLM_KV_HOST, _SHARED and _GIB join the ranks' fingerprint
  
  russellmm
  authored and
  ashhart
  committed
  92e2076
- glm cuda: TF_GLM_KV_HOST's gather kernels join the contract, and a call past its slots is refused
  
  russellmm
  authored and
  ashhart
  committed
  fbd733a
- glm cuda: a TF_GLM_KV_HOST call past its slots is refused by name, with a test
  
  russellmm
  authored and
  ashhart
  committed
  0b4bd44
- changelog: #1795 under Unreleased
  ashhart
  committed
  c0e6daf
- feat(profile): register GLM paging and report resolved selection
  
  garvrao80
  authored and
  ashhart
  committed
  f41bcf4
- changelog: #1959 under Unreleased
  ashhart
  committed
  099ca83
- test(hip): pin the decode softmax stats' max and the partial sums' tile order
  
  AbrahamCete
  authored and
  ashhart
  committed
  f150283
- perf(hip): the decode softmax stats and partial sums load sixteen ahead
  
  AbrahamCete
  authored and
  ashhart
  committed
  74f1625
- changelog: #1961 under Unreleased
  ashhart
  committed
  532589e
- hip flashnext: parallel n-gram gather, release weight pages after upload
  
  gilby
  authored and
  ashhart
  committed
  e0ee513
- zig fmt
  
  gilby
  authored and
  ashhart
  committed
  8f8947e
- changelog: #1611 under Unreleased
  ashhart
  committed
  4a2621a
- rebase: release() takes its scratch at the three callers main gained since #1611
  ashhart
  committed
  104b84a
- rebase: the CUDA load-cache source tests read release()'s new scratch argument
  ashhart
  committed
  2b8d2aa
- build: one core module per XPU tree, passed through runtime(), family() and the module builders
  
  jschmied
  authored and
  ashhart
  committed
  021d9af
- build: the Nemotron XPU device binaries keep a core at optimize, as on main
  
  jschmied
  authored and
  ashhart
  committed
  d80897e
- changelog: #1918 under Unreleased
  ashhart
  committed
  2e56fe1
- cuda: prioritize verified copies in Nemotron lanes
  ashhart
  committed
  a566da5
- changelog: Nemotron CUDA copy drafts under Unreleased
  ashhart
  committed
  071cefe
- flashnext cuda: the MTP head's experts run on the call's last row only
  
  jschmied
  authored and
  ashhart
  committed
  b5c96c5
- changelog: #1925 under Unreleased
  ashhart
  committed
  ad98b60
- glm cuda: a prompt chunk's MTP pass runs its last row past the caches
  
  gilby
  authored and
  ashhart
  committed
  dd326f4
- changelog: #1577 under Unreleased
  ashhart
  committed
  e002bf9
- Account for whole GLM snapshots restored from disk
  ashhart
  committed
  6d8f689
- changelog: GLM on Metal: charge disk-restored prompt states at their full size
  ashhart
  committed
  d38bba3
- test(hip): re-derive every gate-5 record with ROCm 10.1
  
  laudney
  authored and
  ashhart
  committed
  9eb5579
- changelog: #1979 under Unreleased
  ashhart
  committed
  cb972b8
- feat(cli): add prompt-cache-disk alias and startup reporting
  
  garvrao80
  authored and
  ashhart
  committed
  41a2e8c
- fix(cli): reject mixed disk cache spellings and clarify errors
  
  garvrao80
  authored and
  ashhart
  committed
  2730569
- changelog: #1908 under Unreleased
  ashhart
  committed
  3bc3c9e
- fix(flashnext): keep the ranked vtable static
  
  laudney
  authored and
  ashhart
  committed
  95453c5
- feat(cuda): serve --tp to the CUDA families with ranks
  
  laudney
  authored and
  ashhart
  committed
  1e7f5db
- feat(flashnext): ranks agree on a computed checkpoint digest
  
  laudney
  authored and
  ashhart
  committed
  d6b3e6e
- feat(flashnext): accept BF16 dense and shared layers in ModelOpt
  
  laudney
  authored and
  ashhart
  committed
  bfc2cc4
- feat(flashnext): FP8 rows in the n-gram table
  
  laudney
  authored and
  ashhart
  committed
  89a6600
- feat(flashnext): FP8 block experts in the MTP head
  
  laudney
  authored and
  ashhart
  committed
  9b6d865
- docs(flashnext): serve NVIDIA's NVFP4 checkpoint at TP=2
  
  laudney
  authored and
  ashhart
  committed
  7e4c02d
- refactor(cuda): keep --tp's refusals and test out of cuda.zig
  
  laudney
  authored and
  ashhart
  committed
  a1f124b
- test(contract): gate Flash Next NVFP4's BF16 and FP8 expert kernels
  
  laudney
  authored and
  ashhart
  committed
  e516442
- fix(flashnext): keep one GPU's expert sums at --tp 2
  
  laudney
  authored and
  ashhart
  committed
  2676a56
- docs(changelog): note --tp 2 replies equal one GPU's for Flash Next
  
  laudney
  authored and
  ashhart
  committed
  afb8293
- style(contract): shorten comments for the lean rule
  
  laudney
  authored and
  ashhart
  committed
  6afbcdc
- fix(flashnext): word the --tp load refusals and credit the PR
  
  laudney
  authored and
  ashhart
  committed
  00db33f
- changelog: #1771's line once (the rebase's keep-both copied it three times)
  ashhart
  committed
  6194b8c
- cluster: a checkpoint file's header length is checked with core's headerSize before it is read
  ashhart
  committed
  e463344
- qwen_vl: the Metal tower's shard headers are read with core's headerSize, host-tested without Metal
  ashhart
  committed
  e7a4c2b
- changelog: header readers onto core under Unreleased
  ashhart
  committed
  87bf85f
- hip: decode router over the bf16 rows the checkpoint holds
  
  gilby
  authored and
  ashhart
  committed
  2ff4e6e
- contract: the router runner plants a second expert's row only when there is one
  
  gilby
  authored and
  ashhart
  committed
  9c35198
- changelog: #1614 under Unreleased
  ashhart
  committed
  099b2d9
- build: compile-metal-residency-host takes the engine's living module (native/metal_slide.zig imports it)
  ashhart
  committed
  0b10f70
- qwen3_5 cuda: the weight loaders' host bf16 copies are core's bf16_round.bits (#1863)
  ashhart
  committed
  db04139
- changelog: Qwen 3.5 CUDA bf16 copies under Unreleased
  ashhart
  committed
  6703e91
- core: seven host bf16 copies with a plain + round as bits(), negative NaNs no longer overflow
  
  jschmied
  authored and
  ashhart
  committed
  4c75dce
- changelog: #1954 under Unreleased
  ashhart
  committed
  6a50486
- glm cuda: TF_GLM_KDA_PAD keeps the fused KDA input stack packed at TP=3, split untouched
  
  BadAd84
  authored and
  ashhart
  committed
  96b3b7a
- glm cuda: TF_GLM_KDA_PAD - the prompt conv shift strides by the padded width too
  
  BadAd84
  authored and
  ashhart
  committed
  316c0bc
- glm cuda: TF_GLM_KDA_PAD - the shared prompt window strides by the padded width
  
  BadAd84
  authored and
  ashhart
  committed
  842bba3
- glm cuda: TF_GLM_KDA_PAD - the segment layout's row width is padded too
  
  BadAd84
  authored and
  ashhart
  committed
  6f7f99c
- glm cuda: TF_GLM_KDA_PAD pads to ONE width for every rank
  
  BadAd84
  authored and
  ashhart
  committed
  831e20c
- glm cuda: TF_GLM_KDA_PAD - the KDA input linear is packed at the padded width
  
  BadAd84
  authored and
  ashhart
  committed
  b7af934
- glm cuda: KDA pad - one width for every layer and rank, MTP layer included
  
  BadAd84
  authored and
  ashhart
  committed
  6bafb5a
- glm cuda: the pad's comments are one line each (lean check)
  
  BadAd84
  authored and
  ashhart
  committed
  9db24e7
- glm cuda: TF_GLM_KDA_PAD pads a split KDA stack's bf16 rest too
  
  BadAd84
  authored and
  ashhart
  committed
  2d2a84f
- glm cuda: glm-kda-pad-split's header is one line (lean check)
  
  BadAd84
  authored and
  ashhart
  committed
  367e450
- glm cuda: TF_GLM_KDA_PAD read once into Settings, refused without glm_mm, the unpadded stack freed (review)
  
  BadAd84
  authored and
  ashhart
  committed
  c309bb3
- changelog: #1545 under Unreleased
  ashhart
  committed
  6184cee
- docs and GLM config: say where decisions are served, refuse a non-bf16 activation dtype
  ashhart
  committed
  4deaee2
- changelog: GLM activation dtype refusal under Unreleased
  ashhart
  committed
  ef9fa70
- rope: port the three exact frequency copies onto core's helper
  ashhart
  committed
  66924b5
- changelog: #1995 under Unreleased
  ashhart
  committed
  1e11c0e
- glm cuda: the idle KV pool's line reads even under TF_GLM_KV_SHARE
  
  Bizuayeu
  authored and
  ashhart
  committed
  15af823
- changelog: #1993 under Unreleased
  ashhart
  committed
  9eda955
- test(cost cache): count a Windows replace-window miss instead of failing the race
  ashhart
  committed
  f03d8db
- changelog: #2001 under Unreleased
  ashhart
  committed
  824e562
- xpu: the Nemotron loader and the XPU safetensors reader read shard headers with core
  
  jschmied
  authored and
  ashhart
  committed
  fb7ee09
- changelog: #1919 under Unreleased
  ashhart
  committed
  060ddc0
- dsv41 metal: bf16 and fp32 matvecs and the phase-1 glue
  
  gilby
  authored and
  ashhart
  committed
  d8b80bd
- dsv41 metal: the registered glue kernels (norms, RoPE, FP8 and FP4 round trips, embed, join, mean)
  
  gilby
  authored and
  ashhart
  committed
  a23942a
- build: test-info-config runs info's issue 934 test, which its CUDA CLI root never reached
  
  jishnuvenugopal
  authored and
  ashhart
  committed
  168dd12
- changelog: #2006 under Unreleased
  ashhart
  committed
  0f7f5ed
- docs: clarify tool-call scope of cache benchmark receipts
  
  garvrao80
  authored and
  ashhart
  committed
  ed2cca9
- changelog: #1999 under Unreleased
  ashhart
  committed
  b1f8bad
- dsv41 metal: indexer scores, top-512, candidate blocks and the MoE route
  
  gilby
  authored and
  ashhart
  committed
  cca9ccb
- dsv41 metal: the KV compressor and Engram
  
  gilby
  authored and
  ashhart
  committed
  acf12da
- dsv41 metal: hyper-connection mix, pre and post, and the fused and folded decode glue
  
  gilby
  authored and
  ashhart
  committed
  3ef0357
- dsv41 metal: one decoder layer's launch over 1-16 rows, checked stage by stage against float64
  
  gilby
  authored and
  ashhart
  committed
  b85ac2e
- dsv41 metal: the decode forward over every layer, Engram, the head and the DSpark taps
  
  gilby
  authored and
  ashhart
  committed
  7258db3
- dsv41 metal: prompt chunks and the exact decoder suffix
  
  gilby
  authored and
  ashhart
  committed
  e1fd2a2
- dsv41 metal: DSpark drafts
  
  gilby
  authored and
  ashhart
  committed
  d660dbc
- dsv41: the served engine's lane half as a backend-neutral module
  
  gilby
  authored and
  ashhart
  committed
  599fa82
- changelog: #1719 under Unreleased
  ashhart
  committed
  cfcbecb
- server: one-call streams send the whole reply's content (#1396)
  devopsfortunato
  committed
  6c30fd8
- server: GLM tool calls stream their arguments as they are written (#1975)
  devopsfortunato
  committed
  3fda03c

Commits

Commits on Nov 2, 2026

Commits on Nov 3, 2026