Repository navigation

Merge upstream main - #53

Merged

plotarmordev merged 187 commits into

mainfrom

feat/62-merge-upstream-main

Nov 13, 2026

Merged

Merge upstream main#53
plotarmordev merged 187 commits into
mainfrom
feat/62-merge-upstream-main

Commits on Nov 13, 2026
- flashnext cuda: the rank collective's in-place FP32 reduce-scatter, bounded and aborted as its sum
  
  swbiggart
  authored and
  ashhart
  committed
  58bd8ac
- flashnext cuda: a prompt pass at --tp 2 reduce-scatters its experts' slots by rows and gathers their sums
  
  swbiggart
  authored and
  ashhart
  committed
  77b0260
- changelog, recipe and ledger: Flash Next's prompt-pass expert exchange at --tp 2
  
  swbiggart
  authored and
  ashhart
  committed
  119070d
- changelog: #2977's line from lines/2977.txt in place of its own
  ashhart
  committed
  f803fa3
- Remove unused Qwen27 GDN prompt tail scratch
  ashhart
  committed
  40ec17f
- changelog: #2915 under Unreleased
  ashhart
  committed
  47d08dd
- core: Admit and Ledger move to engine_admission.zig, unchanged
  
  kotevcode
  authored and
  ashhart
  committed
  4c0cfb7
- changelog: #2896's line names its pull request
  
  kotevcode
  authored and
  ashhart
  committed
  ff3e811
- changelog: #2896's line from lines/2896.txt in place of its own
  ashhart
  committed
  1ddbd50
- qwen3 cuda: Qwen3 reranker checkpoints on the native CUDA engine
  
  Arminova
  authored and
  ashhart
  committed
  4a21b36
- qwen3 cuda: the pass's arguments in a checked function with a host test; the docs and file headers name the reranker
  
  Arminova
  authored and
  ashhart
  committed
  d5d31e6
- qwen3 cuda: the reranker gate packs a decision at the chunk and block edges the cases' short common prefix never reaches
  
  Arminova
  authored and
  ashhart
  committed
  e879cf5
- changelog: Qwen3 rerankers on CUDA under Unreleased (#2995)
  
  Arminova
  authored and
  ashhart
  committed
  ac81368
- qwen3 cuda: the rerank docs name the checkpoints each native engine serves; the weights header names the head's stored name
  
  Arminova
  authored and
  ashhart
  committed
  e5943ff
- hip flashnext: the MTP head absorbs a prompt chunk on the prompt tiles
  
  gilby
  authored and
  ashhart
  committed
  39946c9
- flashnext hip: the head's run takes the prefill flag as an argument, not from the options
  
  gilby
  authored and
  ashhart
  committed
  d1fecd7
- changelog: the MTP head's prompt-tile line moves to the top of Unreleased
  
  gilby
  authored and
  ashhart
  committed
  c2ee927
- changelog: the MTP head's prompt-tile line carries the cold first-token gain of its receipt
  
  gilby
  authored and
  ashhart
  committed
  52ee805
- dsv41 metal: the lane target, the prompt cache and the native server over Metal
  
  gilby
  authored and
  ashhart
  committed
  f9df8bf
- dsv41 metal: decode plain at four or more active streams, where drafts lose
  
  gilby
  authored and
  ashhart
  committed
  24a1f2f
- dsv41 metal: decode plain from three active streams, not four
  
  gilby
  authored and
  ashhart
  committed
  3f8beec
- dsv41 metal: plain rounds skip the draft pass, and do not launch one for the next
  
  gilby
  authored and
  ashhart
  committed
  f470802
- dsv41 metal: the plain switch line says three streams and the skipped pass
  
  gilby
  authored and
  ashhart
  committed
  8317e6b
- dsv41 metal: a stream's first draft ask waits for its first round
  
  gilby
  authored and
  ashhart
  committed
  cd31156
- dsv41 metal: a stream drafts nothing until its burst has formed
  
  gilby
  authored and
  ashhart
  committed
  873a4bc
- dsv41 metal: restore the engram manifest check, the serve-lane test steps and the forward try's
  
  gilby
  authored and
  ashhart
  committed
  4a5f5d4
- lanes: the plain-streams check in depth.zig on one line, so the file stays within 600 lines
  
  gilby
  authored and
  ashhart
  committed
  c713807
- dsv41 metal: boot and entry read their settings through platform/environment.zig
  
  gilby
  authored and
  ashhart
  committed
  dcd2e9f
- dsv41 metal: TF_DSV41_DRAFT_PLAIN_STREAMS takes 0 to 16 and refuses anything else by name
  
  gilby
  authored and
  ashhart
  committed
  3ee1eab
- dsv41 metal: the plain switch holds a stream's ask only while plain_streams streams decode
  
  gilby
  authored and
  ashhart
  committed
  0c65d23
- dsv41 metal: the taps ring takes a window's kept rows only, so it never holds a rejected draft's taps
  
  gilby
  authored and
  ashhart
  committed
  3f9cd87
- dsv41 metal: plain rounds leave their rows in the target's history, and the next ask catches the drafter up
  
  gilby
  authored and
  ashhart
  committed
  dd60b6e
- dsv41 metal: a round of plain_streams streams verifies no copy windows either
  
  gilby
  authored and
  ashhart
  committed
  1385658
- changelog: #1720's line from lines/1720.txt in place of its own
  ashhart
  committed
  7a96db0
- template: a chat template is held to a nesting depth, and its render to steps, stack and memory
  ashhart
  committed
  3c9fc8e
- template: a step is a statement list or a loop filter's test, and the byte cap counts only what was given
  ashhart
  committed
  def39d2
- changelog: #3072 under Unreleased
  ashhart
  committed
  0591c4e
- slide rollback: a fact graph is removed only when its own folders file names the folder
  ashhart
  committed
  c835f68
- slide rollback: a removed fact graph takes its folders file with it, and a folders file it cannot read leaves the graph
  ashhart
  committed
  8d588d6
- changelog: #3073 under Unreleased
  ashhart
  committed
  13d7ce0
- flashnext hip: a window round's plan buffers
  
  gilby
  authored and
  ashhart
  committed
  07d5fe7
- flashnext hip: a verify window's DeltaNet from its plan, the keep replaying the kept rows
  
  gilby
  authored and
  ashhart
  committed
  18ec707
- flashnext hip: window rounds and their keeps on the plan path
  
  gilby
  authored and
  ashhart
  committed
  43c82e4
- hip kernel tests: Flash Next's plan path against its per-row path, q, k, v, state and conv tail
  
  gilby
  authored and
  ashhart
  committed
  1b2f541
- changelog: Flash Next's plan verify windows under Unreleased
  
  gilby
  authored and
  ashhart
  committed
  c35068d
- test(hip): run Flash Next's conv split with norms and gates on, both paths
  
  gilby
  authored and
  ashhart
  committed
  76e8ea8
- engine(flashnext): free a plan's slots when opening fails, count its bytes once
  
  gilby
  authored and
  ashhart
  committed
  5314474
- test(hip): a Flash Next window through its plan, fill to keep
  
  gilby
  authored and
  ashhart
  committed
  3ae4a10
- engine(flashnext): free a plan's slots when opening fails after them, and keep the kernel-test wiring out of hip.zig
  
  gilby
  authored and
  ashhart
  committed
  bc46621
- changelog: back under Unreleased
  ashhart
  committed
  244a188
- cortheon: keep a page link when a space precedes href's equals
  
  ppardus
  authored and
  ashhart
  committed
  ad0e249
- cortheon: keep a page link when a newline follows href's equals
  
  ppardus
  authored and
  ashhart
  committed
  c4555f7
- cortheon: end an unquoted page link at a tab, CR or LF
  
  ppardus
  authored and
  ashhart
  committed
  3a15de2
- cortheon: resolve a link whose ends are a newline
  
  ppardus
  authored and
  ashhart
  committed
  74beae4
- cortheon: drop a tab or newline wrapped inside a link
  
  ppardus
  authored and
  ashhart
  committed
  9d5258b
- cortheon: keep a page link when an earlier name only contains href
  
  ppardus
  authored and
  ashhart
  committed
  1901362
- changelog: #3006 under Unreleased
  ashhart
  committed
  cc0d6f1
- changelog: #3006 under Unreleased
  ashhart
  committed
  7be6d08
- core: a bounded ladder of intermediate prompt-cache checkpoints
  
  javierbits
  authored and
  ashhart
  committed
  78140ea
- docs: the intermediate checkpoint ladder's GB10 receipt
  
  javierbits
  authored and
  ashhart
  committed
  08265a5
- core: share the interior key check and test the ladder's opt-in
  
  javierbits
  authored and
  ashhart
  committed
  c8cb885
- flashnext: keep cuda_lanes.zig at 600 lines
  
  javierbits
  authored and
  ashhart
  committed
  1d5e7af
- core: ladder checkpoints yield to newest states and to other marks' room
  
  ashhart
  and
  javierbits
  committed
  29b2089
- cuda: read TF_PROMPT_INTERMEDIATE through the strict environment table
  
  ashhart
  and
  javierbits
  committed
  8e98927
- changelog: the ladder's line under Unreleased
  
  ashhart
  and
  javierbits
  committed
  68cde09
- tests: re-pin cuda_lanes.zig and cuda_snap.zig for #1953's checkpoint ladder
  ashhart
  committed
  e3062d2
- changelog: #3006 under Unreleased
  ashhart
  committed
  1a368ec
- qwen27 metal: retain the long-copy driver for a lone greedy stream
  ashhart
  committed
  f986a77
- qwen27 metal: retain the lane head on confident lone copy rounds
  ashhart
  committed
  900d8ea
- changelog: #3069 under Unreleased
  ashhart
  committed
  20c5c61
- changelog: #3006 under Unreleased
  ashhart
  committed
  8e66c92
- changelog: #3069 under Unreleased
  ashhart
  committed
  edd7893
- docs: the CUDA memory plan, what each setting decides and the file that reads it
  
  Bizuayeu
  authored and
  ashhart
  committed
  7e6aec2
- docs: --context's unset and 0 windows, the files that size the unnamed cache, and #3000 for the numbers
  
  Bizuayeu
  authored and
  ashhart
  committed
  0fb14b5
- changelog: #3028 under Unreleased
  ashhart
  committed
  29307ae
- changelog: each line once under Unreleased
  ashhart
  committed
  4a68da2
- changelog: one line for #2728, and #1878's line without its placeholder
  ashhart
  committed
  c34ff56
- qwen27 metal: drop prompt_glue.metal and prompt_state.metal, whose six kernels nothing launches
  
  jschmied
  authored and
  ashhart
  committed
  885918b
- core metal: drop tf_lane_rope and Attention.rotate, which only lanes_see calls
  
  jschmied
  authored and
  ashhart
  committed
  4f12fa0
- runtime sources: drop the Rope loop and rotary fields core.zig still passed to shared_attention; changelog line
  
  jschmied
  authored and
  ashhart
  committed
  e751f29
- deepseek_v41: tf-dsv41-assets builds ReleaseSafe whatever -Doptimize says
  
  ashhart
  and
  Defilan
  committed
  93ad638
- changelog: #3007 under Unreleased
  ashhart
  committed
  d6ddf2b
- cuda experts: a bf16 format, a checkpoint's own weights in mma fragment blocks (relu^2 up, fp32 or bf16 down)
  
  BobClawblaw
  authored and
  ashhart
  committed
  ad3f738
- kernel contract: the bf16 experts (experts_bf16.cu), relu^2 up, fp32 and bf16 down, gates 1-5 with two controls
  
  BobClawblaw
  authored and
  ashhart
  committed
  d0ff0dc
- cuda experts: bf16 run refuses a width or dims off the 32-wide blocks, as pack does, with a test
  
  BobClawblaw
  authored and
  ashhart
  committed
  3033a7c
- tests: re-pin experts.zig and kernels.zig for #1630
  ashhart
  committed
  294f28b
- changelog: #1630 under Unreleased
  ashhart
  committed
  62a9d5e
- glm: the DFlash2 drafter's startup line and docs give its licence as Apache-2.0, as its model card now says
  ashhart
  committed
  c22079b
- changelog: #3084 under Unreleased
  ashhart
  committed
  a3aecd8
- metal: dispatch generic prompt lanes as independent row blocks
  ashhart
  committed
  6667aee
- metal: record measured offline limits for standalone byte-lane sums
  ashhart
  committed
  8c4b12b
- metal: keep short byte-lane projections on a scalar row grid
  ashhart
  committed
  79fc3d4
- metal: opt into whole-prompt row blocks explicitly
  ashhart
  committed
  9359fe7
- changelog: #2851 under Unreleased
  ashhart
  committed
  ae705fe
- glm cuda: TF_GLM_CONTEXT_CEILING serves windows past 1,048,576 tokens, read at GLM's open
  
  Bizuayeu
  authored and
  ashhart
  committed
  36aa830
- glm cuda: TF_GLM_CONTEXT_CEILING's reader joins the .glm check in cuda_env.zig
  
  Bizuayeu
  authored and
  ashhart
  committed
  bbb60d8
- native cuda: the ceiling test and lookup leave cuda.zig for its own test file and cuda_window.zig
  
  Bizuayeu
  authored and
  ashhart
  committed
  fcfa424
- glm cuda: the capacity test leaves cuda_native.zig for cuda_pages_test.zig
  
  Bizuayeu
  authored and
  ashhart
  committed
  90eeb3d
- changelog: #2765 under Unreleased
  ashhart
  committed
  5ef0b1f
- fix(server): let the last --compact-at replace the earlier one
  
  laudney
  authored and
  ashhart
  committed
  100e4a0
- changelog: #2883 under Unreleased
  ashhart
  committed
  d8a2da4
- test: align Qwen27 reply history with newest-state retention
  ashhart
  committed
  60a6656
- changelog: #2889 under Unreleased
  ashhart
  committed
  f13d3de
- ci: run Metal engine host tests in the macOS nightly
  ashhart
  committed
  72b0d70
- changelog: #2891 under Unreleased
  ashhart
  committed
  77c2c1a
- core: one block-FP8 reader and exact scale widening; Flash-Next MTP experts, Drama 3 and Kolibri-1 read through it
  
  jschmied
  authored and
  ashhart
  committed
  1bc80c5
- core fp8: a byte-length refusal names the byte counts; widen's comment says an f16 NaN stays a NaN
  
  jschmied
  authored and
  ashhart
  committed
  40a68f4
- changelog: #2892 under Unreleased
  ashhart
  committed
  b42b92f
- perf(hip): the routed prompt tile takes K-parallel passes for an item of few rows
  
  AbrahamCete
  authored and
  ashhart
  committed
  8721d03
- test(hip): pin the routed tile's cut from table to launch; rowOf checks bounds (#2878 review)
  
  AbrahamCete
  authored and
  ashhart
  committed
  fb85fe7
- changelog: #2878 under Unreleased
  ashhart
  committed
  ab78aef
- glm cuda: a split warm-up that NCCL fails refuses split by name before the rank stops
  
  Bizuayeu
  authored and
  ashhart
  committed
  df79510
- changelog: #2868 under Unreleased
  
  Bizuayeu
  authored and
  ashhart
  committed
  05984ec
- changelog: #2868's line from lines/2868.txt in place of its own
  ashhart
  committed
  1e8b2c3
- refactor(radix): take a Query in the prompt store's begin and keep
  
  laudney
  authored and
  ashhart
  committed
  8d60e1f
- refactor(radix): one store hold a page through a page registry
  
  laudney
  authored and
  ashhart
  committed
  3154901
- feat(radix): page-major prefix groups on the store's path
  
  laudney
  authored and
  ashhart
  committed
  cfc9004
- feat(radix): checkpoint parts made, restored and reclaimed
  
  laudney
  authored and
  ashhart
  committed
  af86722
- feat(radix): a page-only store resumes at any matched page edge
  
  laudney
  authored and
  ashhart
  committed
  51d0124
- fix(radix): count part pages when planning a pass's marks
  
  laudney
  authored and
  ashhart
  committed
  b1fff1f
- test(radix): a legacy save that reclaims never prunes its leaf
  
  laudney
  authored and
  ashhart
  committed
  5ac1801
- changelog: #2875 under Unreleased
  ashhart
  committed
  c91cc24
- qwen27: slot 0 keeps its rewind point through the warm pass, and a background pass's chunks narrow with depth
  
  ThinkOffApp
  authored and
  ashhart
  committed
  63e9557
- qwen27: the warm pass's yield, the rewind carry and the rewind buffer's choice are tested rules
  
  ThinkOffApp
  authored and
  ashhart
  committed
  a94a166
- qwen27: a turn that resumed inside a reply kept as generated makes no new rewind point
  
  ThinkOffApp
  authored and
  ashhart
  committed
  96ea687
- changelog: #2723's line from lines/2723.txt in place of its own
  ashhart
  committed
  698e301
- cuda source check: declare the SDK names main's kernels use
  
  Bizuayeu
  authored and
  ashhart
  committed
  2c9c299
- changelog: #2902 under Unreleased
  ashhart
  committed
  8fda4b7
- native: the CUDA kept-state identity is written once through any hasher, its keys unchanged
  
  kotevcode
  authored and
  ashhart
  committed
  cbf0ceb
- changelog: #2903's line names its pull request
  
  kotevcode
  authored and
  ashhart
  committed
  1885a17
- changelog: #2903's line from lines/2903.txt in place of its own
  ashhart
  committed
  3f8a87a
- docs(xpu): the B70 prefill row predates 1.1.0's precise prompt DeltaNet (#2859)
  ashhart
  committed
  78b8147
- changelog: #2906 under Unreleased
  ashhart
  committed
  ef7d692
- tools: tfvideo_frames.py checks libtfvideo's RGB frames against PyAV's
  
  swbiggart
  authored and
  ashhart
  committed
  cbdc1fe
- video: libtfvideo.so exports only tf_video_*, so FFmpeg's x86_64 asm links
  
  swbiggart
  authored and
  ashhart
  committed
  752aa6b
- ci: the macOS adapters run check-video-exports-linux beside check-video-layout-linux
  
  swbiggart
  authored and
  ashhart
  committed
  1c110c8
- changelog: #2861's line from lines/2861.txt in place of its own
  ashhart
  committed
  bf9b1f1
- refactor(core): move the memory counts to engine_status.zig
  
  laudney
  authored and
  ashhart
  committed
  d83ad8e
- refactor(server): move the JPEG encoder out of pdf.zig
  
  laudney
  authored and
  ashhart
  committed
  6057b7f
- feat(core): add OCR reader types and the image frame size
  
  laudney
  authored and
  ashhart
  committed
  2a0dd79
- feat(server): carry image frames and OCR limits through chat.prepare
  
  laudney
  authored and
  ashhart
  committed
  6ea7144
- feat(ocr): add the shared page steps for the OCR route
  
  laudney
  authored and
  ashhart
  committed
  e34d2eb
- feat(ocr): add the dots.mocr page reader
  
  laudney
  authored and
  ashhart
  committed
  fda30cc
- feat(ocr): POST /v1/ocr in the Mistral OCR contract for images
  
  laudney
  authored and
  ashhart
  committed
  612f120
- docs(ocr): document POST /v1/ocr in the API guide
  
  laudney
  authored and
  ashhart
  committed
  eaaf866
- refactor(ocr): keep later steps' code out of step 1
  
  laudney
  authored and
  ashhart
  committed
  aa8798a
- fix(ocr): bound a page at dots.mocr's pixel limit
  
  laudney
  authored and
  ashhart
  committed
  8a20237
- refactor(ocr): leave later-step code out of step 1
  
  laudney
  authored and
  ashhart
  committed
  288ef93
- test(ocr): refuse an oversized page and check no_vision's text
  
  laudney
  authored and
  ashhart
  committed
  f70b4ef
- docs(dots): name dots.mocr's licences for the prompt and fixtures
  
  laudney
  authored and
  ashhart
  committed
  d3d8c68
- refactor(ocr): drop Source.name until a reader reads it
  
  laudney
  authored and
  ashhart
  committed
  d540a7c
- fix(dots): read with our own prompt and hash-only fixtures
  
  laudney
  authored and
  ashhart
  committed
  02f1ab3
- docs(ocr): restore the OCR section the rebase dropped
  
  laudney
  authored and
  ashhart
  committed
  7187c91
- test(ocr): an image-only engine refuses POST /v1/ocr
  
  laudney
  authored and
  ashhart
  committed
  36b75ee
- changelog: #2570 under Unreleased
  ashhart
  committed
  484437b
- XPU: gdn_out rescales by one multiply in ldexp's single-multiply range, reads row exponents from the bits, and takes e^la once a row
  
  dinadur
  authored and
  ashhart
  committed
  f454f34
- changelog: #2921 under Unreleased
  ashhart
  committed
  aaac299
- ci: a newer push to main cancels main's older run, and the app checks run on pushes to main only
  ashhart
  committed
  44c5654
- changelog: #3230 under Unreleased
  ashhart
  committed
  621308b
- moe_route: router logits for any hidden width a multiple of 1024, and bf16 logits on request
  
  BobClawblaw
  authored and
  ashhart
  committed
  16dfe4e
- flint: serve Recursive Flint (40B-A2.5B MoE) natively on Metal
  
  BobClawblaw
  authored and
  ashhart
  committed
  313d822
- load budget: one core helper for Step-5 and Flint; a limit past the Mac's RAM is refused, not converted
  
  BobClawblaw
  authored and
  ashhart
  committed
  0c3a24d
- docs: Flint's licenses, the weights' and the rec_mlx reference's
  
  BobClawblaw
  authored and
  ashhart
  committed
  f7074cf
- flint: SwiGLU's sigmoid in bf16 steps, as MLX runs it on a bf16 gate (the fp32 form missed MLX on 1,360 of 65,280 bf16 gates, and gave 0 below -17); the router keeps MLX's fp32 form
  
  BobClawblaw
  authored and
  ashhart
  committed
  da10097
- contract: qwen27 attention_io registers through its own Suite, so metal_main has room under 600 lines
  
  BobClawblaw
  authored and
  ashhart
  committed
  a20250e
- kernel contract: Flint's Metal kernels registered (glue, routing, prep, decode and prompt attention; gates 1-4 tests, 16 planted controls, gate-5 records), and core tf_route_logits at Flint's, GLM'…
  
  BobClawblaw
  authored and
  ashhart
  committed
  a5037cb
- kernel contract: Flint's prompt attention counts P at the precision the tensor units keep (fp32 cut to 10 mantissa bits, measured on the device before the gates run), not bf16's
  
  BobClawblaw
  authored and
  ashhart
  committed
  4edbe0d
- changelog: #1177 under Unreleased
  ashhart
  committed
  7d365c3
- qwen35 cuda: skip the one-row rollback copy when nothing keeps it
  ashhart
  committed
  d8d0b4e
- dense serving test: re-pin qwen35.cu to #2933's bytes
  ashhart
  committed
  a62840e
- changelog: #2933 under Unreleased
  ashhart
  committed
  8642b1e
- server: cancel a running or queued request by its id
  ashhart
  committed
  38a339c
- changelog: #3193 under Unreleased
  ashhart
  committed
  1af23a6
- qwen35 cuda: bucket 2-8-row windows at 4 and 8 rows
  ashhart
  committed
  c247ca2
- changelog: #2934 under Unreleased
  ashhart
  committed
  d1dae4f
- video: libtfvideo.so exports only tf_video_*, so FFmpeg's x86_64 asm links
  
  swbiggart
  authored and
  ashhart
  committed
  4931941
- video: libtfvideo decodes H.264 without B-frames and MPEG-4 with frame threads
  
  swbiggart
  authored and
  ashhart
  committed
  0229e2b
- changelog: libtfvideo's frame threads for H.264 without B-frames and MPEG-4, under Unreleased
  
  swbiggart
  authored and
  ashhart
  committed
  57e82eb
- changelog: #2863's line from lines/2863.txt in place of its own
  ashhart
  committed
  577410d
- changelog: back under Unreleased
  ashhart
  committed
  4cfd955
- tests: the Responses cancel test waits for the first token before it cancels
  ashhart
  committed
  a7f67b8
- tests: the dashboard page's test runs on every OS, and expects only tiles the page has
  ashhart
  committed
  d8673e9
- changelog: #2946 under Unreleased
  ashhart
  committed
  b8bebda
- metal: the image towers' bf16 rounding, threadgroup sum and max, and softmax body from one source (#2251)
  
  jschmied
  authored and
  ashhart
  committed
  685bc47
- metal: the image towers' one-rounding LayerNorm statistics and output from one source (#2251)
  
  jschmied
  authored and
  ashhart
  committed
  1301bfd
- changelog: #2948 under Unreleased
  ashhart
  committed
  1e8c7a6
- Merge remote-tracking branch 'upstream/main' into feat/62-merge-upstream-main
  plotarmordev
  committed
  86ca041

Commits

Commits on Nov 13, 2026