FreeToken ROCm Watch — notification channel

[Castoff995]

Этот PR служит постоянным каналом уведомлений для автоматической проверки GitHub каждые 2 часа. Его не нужно ревьюить, мержить или закрывать, пока слежка нужна.

Высший приоритет

• FlashML-org/FreeToken PR feat(rocm): add RDNA3 and RDNA4 runtime foundation FlashML-org/FreeToken#132–feat(rocm): serve on AMD GPUs through the HIP toolchain FlashML-org/FreeToken#137: ROCm/RDNA3/RDNA4 foundation, TVM-FFI HIP JIT, backend gating, RCCL, GGUF RDNA4 и альтернативный ROCm port.
• Issues ROCm RDNA4 gfx1201 Ported. Windows only for now. Test please FlashML-org/FreeToken#82, Add scripts for laying ROCM artifacts out for Python packaging. ROCm/TheRock#120, Uplift how artifacts are handled. ROCm/TheRock#122 и roadmap FreeToken Roadmap (2026) FlashML-org/FreeToken#79.
• Новые upstream issues/PR с AMD, ROCm, HIP, RDNA4, gfx1200, gfx1201, RX 9070, Windows, Triton, TVM-FFI, pinned/mapped host memory, hipHostGetDevicePointer, launch_pdl, clang-cl, MoE/offload.

Рабочие форки

• PialGhosh2233/FreeToken-rocm-gfx1200:gfx1200-moe-gguf.
• Maxritz/FreeToken-rocm-test:main.

Зависимости

• ROCm/TheRock — Windows/RDNA4/gfx1201 wheels, PyTorch/HIP regressions and releases.
• triton-lang/triton-windows — Windows AMD/HIP support and releases.
• apache/tvm-ffi — Windows/HIP JIT and cache/toolchain changes.

Устойчивость модели

Также отслеживаются изменения, способные повлиять на первый токен и практический перевод: Qwen3.6/NVFP4, MoE offload, hybrid_radix, pinned/pageable residency, mapped device pointers, long-prefill/KV-cache hangs and engine crashes.

Уведомление публикуется только при существенном изменении: новый релевантный issue/PR, новый commit в ключевом PR/форке, изменение draft/merge/CI/state, новый релиз либо подтверждённый gfx1201/Windows результат. При отсутствии новостей комментариев не будет.

[github-actions]

@Castoff995 FreeToken ROCm Watch запущен. Проверка выполняется каждые 2 часа; при отсутствии существенных изменений комментариев не будет.

Текущий P0 baseline:

• PR Optimize CI/CD builders ROCm/TheRock#132: open, head 45417560, checks pending — feat(rocm): add RDNA3 and RDNA4 runtime foundation
• PR Do not write out a HIP/VERSION file. ROCm/TheRock#133: open / draft, head 89ad7ab0, checks pending — fix(rocm): make TVM-FFI index and store JIT kernels portable to HIP
• PR First draft of full python packaging. ROCm/TheRock#134: open / draft, head 9a523dc8, checks pending — fix(rocm): gate CUDA-only optional backends on ROCm
• PR The codes building issue of unexpected keyword argument 'follow_symlinks' ROCm/TheRock#135: open / draft, head 9d5e3fc8, checks pending — fix(rocm): route tensor parallel communication through RCCL
• PR How to install AMD GPU driver on a clean machine when using TheRock ROCm/TheRock#136: open / draft, head 49591978, checks pending — feat(rocm): enable native GGUF kernels on RDNA4
• PR Fix Python 3.10 issue ROCm/TheRock#137: open, head 7e1d7421, checks pending — feat(rocm): serve on AMD GPUs through the HIP toolchain

Ключевые issues: FreeToken Roadmap (2026), ROCm RDNA4 gfx1201 Ported. Windows only for now. Test please, Native Windows is classified as pin-uncapped, so ROCm/TheRock#112's per-layer residency never auto-engages, Windows/ROCm on RX 9060 XT (gfx1200): packed GGUF working end-to-end (dense + MoE), root cause of the RDNA4 offload-decode TDR, benchmarks vs llama.cpp.

Дополнительно отслеживаются fork commits, ROCm/TheRock, Triton Windows, TVM-FFI, PyTorch gfx1201, Qwen3.6/NVFP4, mapped host pointers, pin-budget, long-prefill и offload stability.

[github-actions]

{"branches":{"FlashML-org/FreeToken:main":{"branch":"main","date":"2026-10-07T07:15:11Z","message":"feat(deepseek-v41): support text and vision serving (FlashML-org#603)","repo":"FlashML-org/FreeToken","sha":"127c85665b49e39d78a5d09b95a92169b5794eb0","url":"FlashML-org@127c85665b49e39d78a5d09b95a92169b5794eb0"},"Maxritz/FreeToken-rocm-test:main":{"branch":"main","date":"2026-09-20T16:29:33Z","message":"Merge pull request Maxritz#5 from uploadyours/fix/windows-rocm-install-and-qwen35moe-gguf","repo":"Maxritz/FreeToken-rocm-test","sha":"e8545daca33f54069eedd71c8afd8a2ff35380bc","url":"Maxritz@e8545daca33f54069eedd71c8afd8a2ff35380bc"},"PialGhosh2233/FreeToken-rocm-gfx1200:gfx1200-moe-gguf":{"branch":"gfx1200-moe-gguf","date":"2026-08-26T07:51:22Z","message":"docs: Gemma-4-26B-A4B QAT q4_0 GGUF results on gfx1200","repo":"PialGhosh2233/FreeToken-rocm-gfx1200","sha":"a338f49b93e2f098eeb3aa83423c9c5386c53ac4","url":"PialGhosh2233@a338f49b93e2f098eeb3aa83423c9c5386c53ac4"}},"discovered":{"FlashML-org/FreeToken#176%22:%7B%22body_hash%22:%2204e37c143183da8a%22,%22comments%22:4,%22kind%22:%22issue%22,%22number%22:176,%22repo%22:%22FlashML-org/FreeToken%22,%22state%22:%22open%22,%22title%22:%22Feature: Predictive Expert Offloading","updated_at":"2026-10-01T10:46:25Z","url":"https://github.com/FlashML-org/FreeToken/issues/176"},"FlashML-org/FreeToken#239":{"body_hash":"e0d9878137f1a275","comments":3,"kind":"issue","number":239,"repo":"FlashML-org/FreeToken","state":"open","title":"RTX 2050 4GB: Qwen3.6-35B-A3B-NVFP4 runs at ~21 tok/s with single-layer CPU MoE prefill buffer","updated_at":"2026-09-30T11:33:06Z","url":"https://github.com/FlashML-org/FreeToken/issues/239"},"FlashML-org/FreeToken#24":{"body_hash":"04716372ce11a2b6","comments":3,"kind":"pr","number":24,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(cuda): support Turing GPUs","updated_at":"2026-10-01T16:28:42Z","url":"https://github.com/FlashML-org/FreeToken/pull/24"},"FlashML-org/FreeToken#260":{"body_hash":"e0794d434a8a819d","comments":9,"kind":"pr","number":260,"repo":"FlashML-org/FreeToken","state":"open","title":"ROCm: validate native gfx1151 serving on Strix Halo","updated_at":"2026-09-30T15:09:34Z","url":"https://github.com/FlashML-org/FreeToken/pull/260"},"FlashML-org/FreeToken#301":{"body_hash":"40475c88b787fc6e","comments":3,"kind":"issue","number":301,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(server): allow overriding the distributed port","updated_at":"2026-10-06T11:01:11Z","url":"https://github.com/FlashML-org/FreeToken/issues/301"},"FlashML-org/FreeToken#302":{"body_hash":"53ce92fba4ae82bc","comments":4,"kind":"issue","number":302,"repo":"FlashML-org/FreeToken","state":"open","title":"feat(server): add a server-side thinking override","updated_at":"2026-09-29T17:06:47Z","url":"https://github.com/FlashML-org/FreeToken/issues/302"},"FlashML-org/FreeToken#335":{"body_hash":"42a24e63af908edf","comments":1,"kind":"issue","number":335,"repo":"FlashML-org/FreeToken","state":"open","title":"[Bug] b12x NVFP4 backend on sm_120 (RTX 5090) serves batch 1 only: not CUDA-graph capturable, and the worker dies with a tile-config ValueError at batch >= 2 (Qwen3.8-Flash-Next-NVFP4)","updated_at":"2026-10-06T07:11:43Z","url":"https://github.com/FlashML-org/FreeToken/issues/335"},"FlashML-org/FreeToken#338":{"body_hash":"d03533086a4e931f","comments":6,"kind":"pr","number":338,"repo":"FlashML-org/FreeToken","state":"closed","title":"perf(ple): fuse the n-gram row-id hash into one Triton kernel","updated_at":"2026-10-04T11:41:35Z","url":"https://github.com/FlashML-org/FreeToken/pull/338"},"FlashML-org/FreeToken#339":{"body_hash":"7d0f17426b165cb7","comments":4,"kind":"pr","number":339,"repo":"FlashML-org/FreeToken","state":"closed","title":"fix(kernels): drop the D2H sync from the varlen GDN/KDA prefill conv","updated_at":"2026-10-04T11:41:36Z","url":"https://github.com/FlashML-org/FreeToken/pull/339"},"FlashML-org/FreeToken#340":{"body_hash":"4c2febbad30c626f","comments":1,"kind":"pr","number":340,"repo":"FlashML-org/FreeToken","state":"closed","title":"fix(engine): reserve the pool's dummy page in the --moe-cache-auto KV floor","updated_at":"2026-10-04T11:41:36Z","url":"https://github.com/FlashML-org/FreeToken/pull/340"},"FlashML-org/FreeToken#384":{"body_hash":"a5cfef075e1a00c8","comments":3,"kind":"issue","number":384,"repo":"FlashML-org/FreeToken","state":"open","title":"Build on Windows?","updated_at":"2026-09-30T14:31:07Z","url":"https://github.com/FlashML-org/FreeToken/issues/384"},"FlashML-org/FreeToken#390":{"body_hash":"a2503c9f178eed1b","comments":9,"kind":"pr","number":390,"repo":"FlashML-org/FreeToken","state":"closed","title":"fix(models): better support for mixed-precision compressed-tensors NVFP4","updated_at":"2026-10-07T02:20:53Z","url":"https://github.com/FlashML-org/FreeToken/pull/390"},"FlashML-org/FreeToken#399":{"body_hash":"227a0901fd907e75","comments":2,"kind":"pr","number":399,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(cpu-moe): honor padded fp8 scale strides, add avx512f tier and float64 parity (builds on FlashML-org#36)","updated_at":"2026-10-05T06:31:21Z","url":"https://github.com/FlashML-org/FreeToken/pull/399"},"FlashML-org/FreeToken#409":{"body_hash":"ae748f8c8ddac88b","comments":6,"kind":"issue","number":409,"repo":"FlashML-org/FreeToken","state":"open","title":"Qwen3.8-Flash-Next-NVFP4 cannot start on a 12 GB RTX 3060: ~9.9 GiB of unquantized (BF16) attention/embedding weights leave no room for MoE cache + KV","updated_at":"2026-10-06T07:11:46Z","url":"https://github.com/FlashML-org/FreeToken/issues/409"},"FlashML-org/FreeToken#444":{"body_hash":"97cdf976b5f6c886","comments":6,"kind":"issue","number":444,"repo":"FlashML-org/FreeToken","state":"open","title":"[Model] DeepSeek-V4.1-Flash (deepseek_v41) — hybrid CPU-MoE makes it viable on a single GPU, and Engram is already isolated in 2 shards","updated_at":"2026-10-05T06:43:15Z","url":"https://github.com/FlashML-org/FreeToken/issues/444"},"FlashML-org/FreeToken#460":{"body_hash":"efe5429c0d3bc8aa","comments":2,"kind":"pr","number":460,"repo":"FlashML-org/FreeToken","state":"open","title":"feat(models): support deepseek v4.1 with native fp8-fp4 kv storage","updated_at":"2026-10-05T06:50:13Z","url":"https://github.com/FlashML-org/FreeToken/pull/460"},"FlashML-org/FreeToken#534":{"body_hash":"265c914cd1ae1994","comments":4,"kind":"issue","number":534,"repo":"FlashML-org/FreeToken","state":"open","title":"FP8 block-quantized MoE models have no usable kernel under moe_strategy=offload (0.1.3)","updated_at":"2026-10-05T06:32:52Z","url":"https://github.com/FlashML-org/FreeToken/issues/534"},"FlashML-org/FreeToken#538":{"body_hash":"aa31212e0c6a4476","comments":3,"kind":"issue","number":538,"repo":"FlashML-org/FreeToken","state":"open","title":"RTX 3090 device loss (Xid 79) at --max-prefill-length 8192 in offload mode — known limitation, or host-specific?","updated_at":"2026-09-30T14:30:39Z","url":"https://github.com/FlashML-org/FreeToken/issues/538"},"FlashML-org/FreeToken#556":{"body_hash":"78df67c573b9e18a","comments":1,"kind":"issue","number":556,"repo":"FlashML-org/FreeToken","state":"closed","title":"[Desktop] weights failed to load: Qwen3.6-27B NVFP4","updated_at":"2026-10-05T06:28:37Z","url":"https://github.com/FlashML-org/FreeToken/issues/556"},"FlashML-org/FreeToken#561":{"body_hash":"effac85bee1affd0","comments":2,"kind":"issue","number":561,"repo":"FlashML-org/FreeToken","state":"open","title":"Bug: AssertionError: Not enough memory for KV cache when serving Qwen3.8-27B-NVFP4 on Windows","updated_at":"2026-10-02T12:42:09Z","url":"https://github.com/FlashML-org/FreeToken/issues/561"},"FlashML-org/FreeToken#566":{"body_hash":"2c540d2c236623f4","comments":0,"kind":"issue","number":566,"repo":"FlashML-org/FreeToken","state":"closed","title":"QSA block top-k runs a 1M-token window on one program per row","updated_at":"2026-09-29T23:29:18Z","url":"https://github.com/FlashML-org/FreeToken/issues/566"},"FlashML-org/FreeToken#567":{"body_hash":"6af3e9c4ead76444","comments":0,"kind":"issue","number":567,"repo":"FlashML-org/FreeToken","state":"open","title":"A stream that is silent through a long prefill is dropped by Node clients and prefilled for nobody","updated_at":"2026-09-29T17:06:19Z","url":"https://github.com/FlashML-org/FreeToken/issues/567"},"FlashML-org/FreeToken#568":{"body_hash":"f836f522139aa72f","comments":0,"kind":"issue","number":568,"repo":"FlashML-org/FreeToken","state":"open","title":"Serving Qwen3.8-Flash-Next's 1M-token window on one 24 GB card","updated_at":"2026-09-29T17:06:55Z","url":"https://github.com/FlashML-org/FreeToken/issues/568"},"FlashML-org/FreeToken#569":{"body_hash":"194b687889102bba","comments":0,"kind":"issue","number":569,"repo":"FlashML-org/FreeToken","state":"open","title":"A hybrid model's cached prefix cannot be resumed far back: no GDN snapshot survives below a prompt's last chunk","updated_at":"2026-09-29T17:06:23Z","url":"https://github.com/FlashML-org/FreeToken/issues/569"},"FlashML-org/FreeToken#570":{"body_hash":"ccadf4bf679191a5","comments":0,"kind":"issue","number":570,"repo":"FlashML-org/FreeToken","state":"open","title":"On Windows the engine sizes its pools as if no other process held any of the card","updated_at":"2026-09-29T17:06:25Z","url":"https://github.com/FlashML-org/FreeToken/issues/570"},"FlashML-org/FreeToken#571":{"body_hash":"a819c75f8b1f5624","comments":1,"kind":"pr","number":571,"repo":"FlashML-org/FreeToken","state":"closed","title":"perf(qsa): reduce block top-k candidates in levels past 16 splits","updated_at":"2026-09-30T14:30:56Z","url":"https://github.com/FlashML-org/FreeToken/pull/571"},"FlashML-org/FreeToken#572":{"body_hash":"57db1db2add83ffd","comments":0,"kind":"pr","number":572,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(server): keep a silent stream alive and notice a client that left","updated_at":"2026-10-06T10:54:30Z","url":"https://github.com/FlashML-org/FreeToken/pull/572"},"FlashML-org/FreeToken#573":{"body_hash":"d90bc09fa5c00943","comments":3,"kind":"pr","number":573,"repo":"FlashML-org/FreeToken","state":"closed","title":"feat(rotary): size the rope table for YaRN and add --hf-overrides","updated_at":"2026-10-01T07:40:28Z","url":"https://github.com/FlashML-org/FreeToken/pull/573"},"FlashML-org/FreeToken#574":{"body_hash":"eb71c7864dfdb816","comments":0,"kind":"pr","number":574,"repo":"FlashML-org/FreeToken","state":"open","title":"feat(engine): --vram-reserve-mb keeps VRAM free at the largest prefill","updated_at":"2026-10-06T11:00:25Z","url":"https://github.com/FlashML-org/FreeToken/pull/574"},"FlashML-org/FreeToken#575":{"body_hash":"84c166abd171effb","comments":0,"kind":"pr","number":575,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(windows): build the extensions and JIT kernels with MSVC","updated_at":"2026-10-06T11:00:27Z","url":"https://github.com/FlashML-org/FreeToken/pull/575"},"FlashML-org/FreeToken#577":{"body_hash":"fb43cf2047e6f67b","comments":3,"kind":"pr","number":577,"repo":"FlashML-org/FreeToken","state":"open","title":"feat(kvcache): keep prefill-chunk GDN checkpoints in a pinned host bank","updated_at":"2026-10-03T04:38:11Z","url":"https://github.com/FlashML-org/FreeToken/pull/577"},"FlashML-org/FreeToken#578":{"body_hash":"129b1abb2df0c38e","comments":0,"kind":"pr","number":578,"repo":"FlashML-org/FreeToken","state":"open","title":"perf(moe): dequant nvfp4 prefill experts by bit placement and load activations contiguously","updated_at":"2026-10-06T11:00:29Z","url":"https://github.com/FlashML-org/FreeToken/pull/578"},"FlashML-org/FreeToken#579":{"body_hash":"8b996dedc8b58ce6","comments":0,"kind":"issue","number":579,"repo":"FlashML-org/FreeToken","state":"open","title":"An offloaded prefill streams every expert of each layer, however short the chunk","updated_at":"2026-09-30T14:30:36Z","url":"https://github.com/FlashML-org/FreeToken/issues/579"},"FlashML-org/FreeToken#580":{"body_hash":"fbb775a52b247fad","comments":0,"kind":"pr","number":580,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(windows): serve over loopback ZMQ and a selector event loop","updated_at":"2026-10-06T11:00:32Z","url":"https://github.com/FlashML-org/FreeToken/pull/580"},"FlashML-org/FreeToken#581":{"body_hash":"31d2cefc49b9a275","comments":0,"kind":"pr","number":581,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(windows): read checkpoints without O_DIRECT, posix_fadvise or PROT_READ","updated_at":"2026-10-06T10:54:53Z","url":"https://github.com/FlashML-org/FreeToken/pull/581"},"FlashML-org/FreeToken#583":{"body_hash":"bb0f92f655d6f410","comments":0,"kind":"pr","number":583,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(engine): size memory pools against the VRAM other processes leave free on Windows","updated_at":"2026-10-06T11:00:35Z","url":"https://github.com/FlashML-org/FreeToken/pull/583"},"FlashML-org/FreeToken#584":{"body_hash":"00499fe994d84ba6","comments":0,"kind":"issue","number":584,"repo":"FlashML-org/FreeToken","state":"open","title":"The computer on the local area network cannot be accessed","updated_at":"2026-09-30T15:54:22Z","url":"https://github.com/FlashML-org/FreeToken/issues/584"},"FlashML-org/FreeToken#585":{"body_hash":"bc29701a8b518529","comments":0,"kind":"pr","number":585,"repo":"FlashML-org/FreeToken","state":"open","title":"perf(linear): run a single-row bf16 projection as one GEMV","updated_at":"2026-10-06T10:54:46Z","url":"https://github.com/FlashML-org/FreeToken/pull/585"},"FlashML-org/FreeToken#587":{"body_hash":"7e6929aa54eb0c69","comments":1,"kind":"pr","number":587,"repo":"FlashML-org/FreeToken","state":"open","title":"feat(rocm): run on RDNA3/RDNA4 with torch 2.14/2.15 and ROCm 7.14/10","updated_at":"2026-10-05T16:16:21Z","url":"https://github.com/FlashML-org/FreeToken/pull/587"},"FlashML-org/FreeToken#588":{"body_hash":"611f86536843cd1d","comments":0,"kind":"pr","number":588,"repo":"FlashML-org/FreeToken","state":"closed","title":"fix(ple): fence disk staging against in-flight lookups","updated_at":"2026-10-01T23:59:45Z","url":"https://github.com/FlashML-org/FreeToken/pull/588"},"FlashML-org/FreeToken#589":{"body_hash":"3545d2d7d7856bb6","comments":0,"kind":"pr","number":589,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(checkpoint): resolve HF repo Ids before conversion","updated_at":"2026-10-02T08:52:03Z","url":"https://github.com/FlashML-org/FreeToken/pull/589"},"FlashML-org/FreeToken#590":{"body_hash":"05469569baac50e7","comments":0,"kind":"pr","number":590,"repo":"FlashML-org/FreeToken","state":"open","title":"perf(model): fuse hc combine with the next block's RMSNorm","updated_at":"2026-10-02T22:44:15Z","url":"https://github.com/FlashML-org/FreeToken/pull/590"},"FlashML-org/FreeToken#594":{"body_hash":"b03118a468b33c59","comments":0,"kind":"pr","number":594,"repo":"FlashML-org/FreeToken","state":"open","title":"feat(qwen4_exp): support ModelOpt QAD checkpoints (NVFP4 PLE tables, MXFP8 dense)","updated_at":"2026-10-03T16:27:55Z","url":"https://github.com/FlashML-org/FreeToken/pull/594"},"FlashML-org/FreeToken#595":{"body_hash":"f619a84581dec764","comments":1,"kind":"pr","number":595,"repo":"FlashML-org/FreeToken","state":"open","title":"feat(server): add ft info and a pre-load memory preflight","updated_at":"2026-10-05T15:37:08Z","url":"https://github.com/FlashML-org/FreeToken/pull/595"},"FlashML-org/FreeToken#596":{"body_hash":"24b00ae8fdb7ece2","comments":0,"kind":"pr","number":596,"repo":"FlashML-org/FreeToken","state":"open","title":"feat(engine): --embed-device cpu keeps the input embedding in RAM","updated_at":"2026-10-04T01:34:09Z","url":"https://github.com/FlashML-org/FreeToken/pull/596"},"FlashML-org/FreeToken#597":{"body_hash":"c34a9a260da47f5f","comments":2,"kind":"issue","number":597,"repo":"FlashML-org/FreeToken","state":"open","title":"[Windows][Desktop] WeightLoadError: invalid python storage with Qwen3.6 NVFP4 and gpt-oss-20b MXFP4","updated_at":"2026-10-05T18:24:01Z","url":"https://github.com/FlashML-org/FreeToken/issues/597"},"FlashML-org/FreeToken#601":{"body_hash":"fdf19066450ceb13","comments":0,"kind":"pr","number":601,"repo":"FlashML-org/FreeToken","state":"closed","title":"feat(moe): serve NVFP4 and MXFP4 experts fused through the expert banks","updated_at":"2026-10-04T20:51:56Z","url":"https://github.com/FlashML-org/FreeToken/pull/601"},"FlashML-org/FreeToken#605":{"body_hash":"f03f9e77e023cc60","comments":0,"kind":"pr","number":605,"repo":"FlashML-org/FreeToken","state":"open","title":"perf(pre-sm70): pipeline the block-fp8 GEMMs and halve the extend-attention K/V re-reads on Pascal","updated_at":"2026-10-05T16:31:10Z","url":"https://github.com/FlashML-org/FreeToken/pull/605"},"FlashML-org/FreeToken#606":{"body_hash":"12d688e163f0e273","comments":1,"kind":"issue","number":606,"repo":"FlashML-org/FreeToken","state":"open","title":"DeepSeek-V4-Flash-0731-NVFP4 (nvidia) fails on 0.1.3+gfa0e9537d: WeightLoadError KeyError 'layers.0.ffn.experts.0.w1.scale'","updated_at":"2026-10-06T02:19:42Z","url":"https://github.com/FlashML-org/FreeToken/issues/606"},"FlashML-org/FreeToken#607":{"body_hash":"5800e5a7dd83d2d0","comments":1,"kind":"issue","number":607,"repo":"FlashML-org/FreeToken","state":"open","title":"Windows wheels resolve unpinned torch dependency to CPU-only build: 'cannot use CUDA device 0: only 0 device(s) visible'","updated_at":"2026-10-05T22:53:59Z","url":"https://github.com/FlashML-org/FreeToken/issues/607"},"FlashML-org/FreeToken#608":{"body_hash":"18c66c42155be45a","comments":0,"kind":"pr","number":608,"repo":"FlashML-org/FreeToken","state":"closed","title":"fix(server): keep Anthropic inline system messages in place for prefix reuse","updated_at":"2026-10-05T22:00:00Z","url":"https://github.com/FlashML-org/FreeToken/pull/608"},"FlashML-org/FreeToken#609":{"body_hash":"071a4e2fb7588698","comments":0,"kind":"pr","number":609,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(models): preserve glm-5.2 router correction bias in fp32","updated_at":"2026-10-05T22:57:35Z","url":"https://github.com/FlashML-org/FreeToken/pull/609"},"FlashML-org/FreeToken#610":{"body_hash":"a9262becc4676eab","comments":5,"kind":"pr","number":610,"repo":"FlashML-org/FreeToken","state":"closed","title":"perf(moe): Improved fast_index_copy_multi_jit for NVLink-C2C hosts (GH200: 222 → 412 GB/s)","updated_at":"2026-10-07T02:54:39Z","url":"https://github.com/FlashML-org/FreeToken/pull/610"},"FlashML-org/FreeToken#611":{"body_hash":"b5243c0b22682524","comments":0,"kind":"issue","number":611,"repo":"FlashML-org/FreeToken","state":"open","title":"GLM-5.3-Flash-NVFP4 CUDA OOM during load on RTX 5070 Ti 16 GB (sm_120)","updated_at":"2026-10-06T07:11:57Z","url":"https://github.com/FlashML-org/FreeToken/issues/611"},"FlashML-org/FreeToken#612":{"body_hash":"44f6d4e3229a047d","comments":0,"kind":"pr","number":612,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(moe): route moe_align to the atomic-free staged path below sm_70","updated_at":"2026-10-06T10:41:03Z","url":"https://github.com/FlashML-org/FreeToken/pull/612"},"FlashML-org/FreeToken#613":{"body_hash":"e09232c609ed4deb","comments":0,"kind":"pr","number":613,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(server)!: answer /health with 503 until the server accepts requests","updated_at":"2026-10-06T11:00:46Z","url":"https://github.com/FlashML-org/FreeToken/pull/613"},"FlashML-org/FreeToken#617":{"body_hash":"b3c1a5f65561a2c8","comments":0,"kind":"pr","number":617,"repo":"FlashML-org/FreeToken","state":"open","title":"perf(fla): fix the pre-sm70 GDN chunk-kernel launch configs (7.4x / 11.3x)","updated_at":"2026-10-06T17:34:53Z","url":"https://github.com/FlashML-org/FreeToken/pull/617"},"FlashML-org/FreeToken#618":{"body_hash":"6704be182c9af027","comments":0,"kind":"pr","number":618,"repo":"FlashML-org/FreeToken","state":"open","title":"feat(attention): hand-written GQA-fused extend kernel for pre-sm70 (1.8x)","updated_at":"2026-10-06T17:35:09Z","url":"https://github.com/FlashML-org/FreeToken/pull/618"},"FlashML-org/FreeToken#619":{"body_hash":"5afe0636b926dc33","comments":2,"kind":"pr","number":619,"repo":"FlashML-org/FreeToken","state":"closed","title":"perf(moe): spin-wait the cpu moe pool between layers","updated_at":"2026-10-07T04:38:11Z","url":"https://github.com/FlashML-org/FreeToken/pull/619"}},"errors":["dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 4845:1683C0:FBE78:16CDA3:6AC60612 and timestamp 2026-10-07 08:42:58 UTC. For more on scraping GitHub and how it may affect your rights, please review our Terms of Service (https://docs.github.com/en/site-policy/github-terms/github-terms-of-service)"],"external_tracked_issues":{"ROCm/legacy-rocm-build#6461":{"body_hash":"0ac6e80ee9b489cd","comments":6,"kind":"issue","number":6461,"relevance":"high","relevance_reason":"Windows + gfx1201 + hipBLASLt/rocBLAS sequence/state-dependent GEMM failure","repo":"ROCm/legacy-rocm-build","state":"closed","title":"[Issue]: rocBLAS bf16/f16 GEMM fails with hipErrorInvalidValue on 2nd call, gfx1201 / Windows / ROCm 7.14","updated_at":"2026-09-23T17:39:44Z","url":"https://github.com/ROCm/legacy-rocm-build/issues/6461"}},"last_checked":"2026-10-07T08:42:27+00:00","releases":{"FlashML-org/FreeToken":{"draft":false,"name":"v0.1.3","prerelease":false,"published_at":"2026-09-16T01:02:33Z","tag":"v0.1.3","url":"https://github.com/FlashML-org/FreeToken/releases/tag/v0.1.3"},"ROCm/ROCm":{"draft":false,"name":"ROCm 7.14.0 Release","prerelease":false,"published_at":"2026-07-16T03:28:58Z","tag":"rocm-7.14.0","url":"https://github.com/ROCm/legacy-rocm-build/releases/tag/rocm-7.14.0"},"ROCm/TheRock":{"draft":false,"name":"","prerelease":false,"published_at":"2026-10-05T10:42:56Z","tag":"therock-10.1","url":"https://github.com/ROCm/TheRock/releases/tag/therock-10.1"},"apache/tvm-ffi":{"draft":false,"name":"v0.1.14-post1","prerelease":false,"published_at":"2026-09-22T12:36:48Z","tag":"v0.1.14-post1","url":"https://github.com/apache/tvm-ffi/releases/tag/v0.1.14-post1"},"ggml-org/llama.cpp":{"draft":false,"name":"v0.6.0","prerelease":false,"published_at":"2026-10-05T16:56:22Z","tag":"v0.6.0","url":"https://github.com/ggml-org/llama.cpp/releases/tag/v0.6.0"},"triton-lang/triton-windows":{"draft":false,"name":"v3.8.0-windows.post29","prerelease":false,"published_at":"2026-09-28T22:00:36Z","tag":"v3.8.0-windows.post29","url":"https://github.com/triton-lang/triton-windows/releases/tag/v3.8.0-windows.post29"}},"tracked_issues":{"110":{"body_hash":"d6672f16a90f2854","comments":3,"kind":"issue","number":110,"repo":"FlashML-org/FreeToken","state":"open","title":"Unhandled CUDA OOM in prefill expert workspace kills the engine; server keeps accepting requests that never return (root cause behind FlashML-org#20-style headless state; likely also FlashML-org#72)","updated_at":"2026-10-06T10:03:58Z","url":"https://github.com/FlashML-org/FreeToken/issues/110"},"111":{"body_hash":"b01b977d143435c7","comments":3,"kind":"issue","number":111,"repo":"FlashML-org/FreeToken","state":"open","title":"Requests longer than the KV pool are queued forever with no error — and --moe-cache-auto leaves only ~8k tokens of KV on a 96GB GPU","updated_at":"2026-08-31T02:09:35Z","url":"https://github.com/FlashML-org/FreeToken/issues/111"},"120":{"body_hash":"a7caa8c0bca18470","comments":1,"kind":"issue","number":120,"repo":"FlashML-org/FreeToken","state":"open","title":"Native Windows is classified as pin-uncapped, so FlashML-org#112's per-layer residency never auto-engages","updated_at":"2026-08-24T03:04:21Z","url":"https://github.com/FlashML-org/FreeToken/issues/120"},"122":{"body_hash":"4d42515b782747ed","comments":6,"kind":"issue","number":122,"repo":"FlashML-org/FreeToken","state":"open","title":"Windows/ROCm on RX 9060 XT (gfx1200): packed GGUF working end-to-end (dense + MoE), root cause of the RDNA4 offload-decode TDR, benchmarks vs llama.cpp","updated_at":"2026-08-27T07:18:05Z","url":"https://github.com/FlashML-org/FreeToken/issues/122"},"123":{"body_hash":"09dd1b9b8cf6560f","comments":1,"kind":"issue","number":123,"repo":"FlashML-org/FreeToken","state":"closed","title":"Request hang (silent, zero-log) with hybrid_radix cache + moe-backend offload + nvfp4 triton backend — reproduced on 2 independent models, not explained by FlashML-org#110","updated_at":"2026-08-29T05:13:10Z","url":"https://github.com/FlashML-org/FreeToken/issues/123"},"124":{"body_hash":"64a1de8500f135a3","comments":0,"kind":"issue","number":124,"repo":"FlashML-org/FreeToken","state":"closed","title":"NVFP4 checkpoints fail to load: model.safetensors.index.json is not downloaded, but the NVFP4 bank loader requires it","updated_at":"2026-09-02T02:43:20Z","url":"https://github.com/FlashML-org/FreeToken/issues/124"},"79":{"body_hash":"1bc2e650d72ca13e","comments":17,"kind":"issue","number":79,"repo":"FlashML-org/FreeToken","state":"open","title":"FreeToken Roadmap (2026)","updated_at":"2026-10-06T17:57:47Z","url":"https://github.com/FlashML-org/FreeToken/issues/79"},"82":{"body_hash":"ff59a7ff00722b01","comments":19,"kind":"issue","number":82,"repo":"FlashML-org/FreeToken","state":"open","title":"ROCm RDNA4 gfx1201 Ported. Windows only for now. Test please","updated_at":"2026-09-15T15:50:45Z","url":"https://github.com/FlashML-org/FreeToken/issues/82"}},"tracked_prs":{"112":{"base_sha":"f0abe587a11cca53bb3c37a9596fad24973ace62","body_hash":"aa539235301fea3d","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":1,"commits":3,"draft":false,"head_sha":"831d38a66bc45543411078c6805de8c42e7603fa","merged":true,"number":112,"review_comments":0,"state":"closed","title":"feat(moe): per-layer host-bank residency","updated_at":"2026-08-24T00:39:32Z","url":"https://github.com/FlashML-org/FreeToken/pull/112"},"118":{"base_sha":"e0a3bbc04652ec0622d932adcbf2772848e3c2e2","body_hash":"d73a76c00cdb2f94","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":0,"commits":1,"draft":false,"head_sha":"6241178a7550e5a0ffadd85f0eb04af0160f9988","merged":false,"number":118,"review_comments":0,"state":"open","title":"fix(scheduler): clamp admission to the actual KV pool so oversized prompts fail loudly","updated_at":"2026-08-24T00:52:33Z","url":"https://github.com/FlashML-org/FreeToken/pull/118"},"125":{"base_sha":"bd372b630a028e3faa51f4ab0ef6a98c2f2de501","body_hash":"793491958f083651","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":0,"commits":1,"draft":false,"head_sha":"cb44bdeedb7ea6ed2b37c5c49864ad40065e1985","merged":false,"number":125,"review_comments":0,"state":"open","title":"test(pinned): use mapped memory for UVA pointer check","updated_at":"2026-08-24T05:48:44Z","url":"https://github.com/FlashML-org/FreeToken/pull/125"},"129":{"base_sha":"bd372b630a028e3faa51f4ab0ef6a98c2f2de501","body_hash":"d5c4ac29d87557fc","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":3,"commits":1,"draft":false,"head_sha":"d30726e006b2708d4fbfead120985bc23c935541","merged":false,"number":129,"review_comments":0,"state":"closed","title":"fix(download): include json files when fetching weights, so NVFP4's index.json is no longer silently skipped","updated_at":"2026-09-02T02:45:39Z","url":"https://github.com/FlashML-org/FreeToken/pull/129"},"131":{"base_sha":"bd372b630a028e3faa51f4ab0ef6a98c2f2de501","body_hash":"33bc52d171e8ca00","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":8,"commits":36,"draft":false,"head_sha":"bb432e8c473c557cdef6517abccf0c129bd8c544","merged":false,"number":131,"review_comments":0,"state":"open","title":"GGUF: support all quant types, add qwen35moe","updated_at":"2026-08-29T22:27:37Z","url":"https://github.com/FlashML-org/FreeToken/pull/131"},"132":{"base_sha":"cc1f5c2c91855f2cc7787ad6b909f7e46a5d5825","body_hash":"f5b92547538a8571","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":8,"commits":11,"draft":false,"head_sha":"c6be4afb541b34f36fb811cd2f6efc01784e9e5d","merged":true,"number":132,"review_comments":7,"state":"closed","title":"feat(rocm): add RDNA3 and RDNA4 runtime foundation","updated_at":"2026-09-26T00:49:46Z","url":"https://github.com/FlashML-org/FreeToken/pull/132"},"133":{"base_sha":"cc1f5c2c91855f2cc7787ad6b909f7e46a5d5825","body_hash":"eb17018abe601f2c","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":1,"commits":11,"draft":true,"head_sha":"9ec8271ab0b824246944f7f23c9e819185ac44b3","merged":false,"number":133,"review_comments":0,"state":"open","title":"fix(rocm): make TVM-FFI index and store JIT kernels portable to HIP","updated_at":"2026-09-21T03:01:45Z","url":"https://github.com/FlashML-org/FreeToken/pull/133"},"134":{"base_sha":"cc1f5c2c91855f2cc7787ad6b909f7e46a5d5825","body_hash":"467c9a375cece285","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":1,"commits":11,"draft":true,"head_sha":"c4f4a36e945a7c13f4f5b500eb2e6a8e07183d4c","merged":false,"number":134,"review_comments":0,"state":"open","title":"fix(rocm): gate CUDA-only optional backends on ROCm","updated_at":"2026-09-21T03:01:49Z","url":"https://github.com/FlashML-org/FreeToken/pull/134"},"135":{"base_sha":"cc1f5c2c91855f2cc7787ad6b909f7e46a5d5825","body_hash":"3a0be6c359e61828","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":5,"commits":12,"draft":true,"head_sha":"3675d03e17110adfcaa532cc965e5c79895cc8f3","merged":false,"number":135,"review_comments":0,"state":"open","title":"fix(rocm): route tensor parallel communication through RCCL","updated_at":"2026-10-06T10:49:08Z","url":"https://github.com/FlashML-org/FreeToken/pull/135"},"136":{"base_sha":"cc1f5c2c91855f2cc7787ad6b909f7e46a5d5825","body_hash":"c875bbbd1cb60d21","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":5,"commits":13,"draft":true,"head_sha":"76bf2ac398464f8762148356b4a4019e3327b2a9","merged":false,"number":136,"review_comments":0,"state":"closed","title":"feat(rocm): enable native GGUF kernels on ROCm (validated on RDNA4)","updated_at":"2026-09-24T06:02:29Z","url":"https://github.com/FlashML-org/FreeToken/pull/136"},"137":{"base_sha":"f7c31e92dbf296de3a5bb1b9c5fcb58f988e3a07","body_hash":"a0597a23e4607628","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":0,"commits":1,"draft":false,"head_sha":"5eec2582ce89376e3f5016d47ff73ca1b4e8d445","merged":false,"number":137,"review_comments":0,"state":"open","title":"feat(rocm): serve on AMD GPUs through the HIP toolchain","updated_at":"2026-08-26T09:03:28Z","url":"https://github.com/FlashML-org/FreeToken/pull/137"},"27":{"base_sha":"0ab982f10905fa775962a4eddcb44caa50065251","body_hash":"e85637efb5176c06","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":4,"commits":1,"draft":false,"head_sha":"d76d5e0105235e9f13edd534fec3dfa86f25cb15","merged":false,"number":27,"review_comments":0,"state":"closed","title":"feat(moe): partial-pin serving to beat the Windows/WSL2 pinned-memory quota (--pin-exempt-layers)","updated_at":"2026-08-24T01:28:09Z","url":"https://github.com/FlashML-org/FreeToken/pull/27"}},"version":1}

Technical baseline updated automatically: 2026-10-07T08:42:27+00:00. This comment is edited in place and does not represent a notification.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-24T15:02:31+00:00).

• [Issue] Windows gfx120X-all Test Artifacts: hipInfo.exe hangs / GPU not detected → cascade failure across 27 jobs: обновлён (state open→open, comments 5→6).
• Новый релевантный issue: bench_offload_cache_copy.py default sweep aborts on minimax-m2.5-marlin and skips the last four profiles.
• Новый релевантный issue: [XPU] inductor workloads can fail intermittently with UR_RESULT_ERROR_PROFILING_INFO_NOT_AVAILABLE (error 17).
• [BUG] preparing weights when converting: обновлён (state open→open, comments 3→4).

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-24T16:56:53+00:00).

• fix(download): include json files when fetching weights, so NVFP4's index.json is no longer silently skipped: обновлены описание/review/discussion (comments 0→1, review 0→0).
• Новый релевантный issue: AssertionError: cache budget too small.
• torch.nn.LSTM.forward produces incorrect outputs (or crashes) on ROCm: обновлён (state open→open, comments 13→14).
• bench_offload_cache_copy.py default sweep aborts on minimax-m2.5-marlin and skips the last four profiles: обновлён (state open→open, comments 0→1).
• feat(server): Add OpenAI /v1/models/{model_id} endpoint: обновлён (state open→open, comments 1→2).
• [Issue]: Windows: rocm-sdk silently drops device files from the devel tree when the link target exceeds MAX_PATH: обновлён (state open→closed, comments 0→0).

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-24T18:31:06+00:00).

• fix(download): include json files when fetching weights, so NVFP4's index.json is no longer silently skipped: обновлены описание/review/discussion (comments 1→2, review 0→0).
• Новый релевантный pr: fix(bench): skip an unsupported slot count instead of aborting the sweep.
• Новый релевантный issue: Eval bug: Windows HIP gfx1201: Qwen3.6-35B-A3B NVFP4 loads, then first MUL_MAT fails with ROCm invalid argument.
• bench_offload_cache_copy.py default sweep aborts on minimax-m2.5-marlin and skips the last four profiles: обновлён (state open→open, comments 1→5).
• Новый релевантный pr: fix(kernel): clear the latched CUDA error before raising in pinned_tensor.
• Новый релевантный pr: CUDA: size routed MoE MMQ N-tiles from typical expert width on RDNA3.
• Новый релевантный issue: Eval bug: inline </think> tag contained in thinking block gets mis-parsed as the end of reasoning content.
• Новый релевантный issue: Misc. bug: RPC + Metal: main node retains the full model with default mmap load mode, capping split models to a single node's memory.
• Новый релевантный pr: AVX2: Speed up large batch size prompt processing of IQ models.
• Новый релевантный pr: CUDA: extend MOE fusion to specdec, earlier MOE glu fusion and topk-router fusion were restricted to 1 token.
• Новый релевантный issue: Feature Request: server in Router Mode: Multiple presets for the same model (loaded once).
• Новый релевантный issue: SYCL: MoE expert tensors never get the Q8_0 reorder path (opt_for_reorder_id).
• Новый релевантный issue: HIP/ROCm backend produces corrupted output on gfx1151: two dense architectures fail, Vulkan is correct with byte-identical flags.
• Новый релевантный pr: CUDA: XOR swizzle flash attn K,V smem fp16 tiles.
• Новый релевантный pr: CUDA: Update Q4_K and Q5_K to use branchless computation, which stops the scale unpack being re-executed for every column in mmvq, improving perf at batch sizes > 1.
• Новый релевантный pr: vulkan: add f16 B-type matmul pipelines and warp tile size tuning for Intel coopmat1.
• Новый релевантный issue: Vulkan: batched decode throughput cliff at n_tokens=9 on many-expert MoE (fixed 8-token thresholds in MMV dispatch).
• Новый релевантный issue: Eval bug: Unexpected massive performace drop with Vulkan Flash Attention.
• Новый релевантный pr: mamba2 : Flatten in/out projections to dispatch GEMM instead of GEMV.
• Новый релевантный issue: Perf: CUDA: Mamba2 batched decode dispatch GEMV instead of GEMM.
• Новый релевантный pr: hexagon: add alternative Hexagon NPU backend implementation.
• Новый релевантный pr: tests: add mul_mat, flash-attention and copy shapes that real models emit.
• Новый релевантный pr: CUDA: Add experimental SM120 CUTLASS MoE prefill for MXFP4 and NVFP4.
• Новый релевантный issue: Misc. bug: Q6_K MUL_MAT produces wrong results on Adreno 740 OpenCL - test-backend-ops fails.
• Новый релевантный issue: Documented Linux HIP Compile Command Fails with pip-installed ROCm.
• Новый релевантный pr: opencl: use right-sized matmul kernels for speculative decoding's verification step.
• Новый релевантный issue: [CUDA][Blackwell] Qwen3.8-27B-NVFP4 decode hangs, CPU spin, no GPU work.
• Новый релевантный pr: metal : skinny-batch mul_mm tiles + deterministic split-K for small-batch decode.
• Новый релевантный issue: Vulkan dense prefill on gfx1151 (RDNA3 iGPU): ~1.8x from a larger K-quant int-mmq tile.
• Новый релевантный issue: Eval bug: MMVQ host/device nwarps mismatch causes NaN on gfx11-generic / RDNA3.5.
• Новый релевантный pr: opencl: tune the quant paths for Intel Xe-LP GPUs to improve its TG and PP performance.
• Новый релевантный pr: opencl: build one program per kernel where the compiler cannot hold several (for Adreno A6x).
• Новый релевантный pr: convert: fix get block count error for Nemotron 3 Ultra.
• Новый релевантный pr: Quant: OCP FP8 E4M3 support.
• Новый релевантный pr: ggml: add support for MXFP8 CPU.
• …ещё 11 изменений скрыто, чтобы комментарий не превращался в свалку.

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-24T19:01:35+00:00).

• CUDA: size routed MoE MMQ N-tiles from typical expert width on RDNA3: обновлён (state open→open, comments 12→14).
• CUDA: XOR swizzle flash attn K,V smem fp16 tiles: обновлён (state open→open, comments 21→22).
• [BUG] preparing weights when converting: обновлён (state open→open, comments 4→5).

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-24T20:49:53+00:00).

• GGUF: support all quant types, add qwen35moe: новый commit 0f26f7ec → 6f6c8640.
• FreeToken Roadmap (2026): comments 2→3.
• Windows/ROCm on RX 9060 XT (gfx1200): packed GGUF working end-to-end (dense + MoE), root cause of the RDNA4 offload-decode TDR, benchmarks vs llama.cpp: comments 0→1.
• Новый релевантный pr: [ROCm][Inductor] Add gfx950 FlexAttention backward default configs.
• Новый релевантный pr: quantization: complete MXFP4 and NVFP4.
• feat(ci): support runtime-only host-ASAN presubmit builds and tests: обновлён (state open→open, comments 4→5).
• Новый релевантный issue: [RFC] Transition ROCm nightly wheels to wheel-based ROCm install (7.14+).
• CUDA: extend MOE fusion to specdec, earlier MOE glu fusion and topk-router fusion were restricted to 1 token: обновлён (state open→open, comments 2→3).
• Новый релевантный pr: Fix hipify skipping sources under Windows subst drives/symlinks in CUDAExtension.

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-24T22:45:58+00:00).

• Новый релевантный pr: [ROCm][c10d] Enable NCCL2 windows and enforce allocator provenance.
• Новый релевантный pr: [ROCm][Inductor] Apply tt.pointer_range guards to template kernels.
• [ROCm][Inductor] Add gfx950 FlexAttention backward default configs: обновлён (state open→open, comments 6→8).
• Новый релевантный issue: Request: RX 6000 (RDNA2) Windows support - native runtime proven (gfx1030/1031/1032).
• Новый релевантный issue: DISABLED test_workspace_runtime_integration (main.TestNVUniversalGemmEpilogueFusion).
• Новый релевантный issue: Please help support older hardware (1080/2080 series).

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T01:53:38+00:00).

• FreeToken Roadmap (2026): comments 3→4.
• [ROCm] Fix hipify skipping sources under Windows subst drives/symlinks in CUDAExtension: обновлён (state open→open, comments 3→3).
• [ROCm][c10d] Enable NCCL2 windows and enforce allocator provenance: обновлён (state open→open, comments 3→3).
• Новый релевантный issue: Eval bug: Vulkan regression on AMD Strix Halo UMA between b9660 and b9672: model loads on GPU but inference shows no GPU compute.
• Новый релевантный issue: Eval bug: CUDA error: unsupported value or parameter in cublasSgemm_v2 during large context processing.
• Новый релевантный issue: Eval bug: Starting from B9992, long contexts trigger a CUDA error: the function failed to launch on the GPU.
• feat(server)!: add --gpu to choose the GPU on multi-GPU machines: обновлён (state open→closed, comments 0→1).
• Difficult to load model despite more than enough vram/ram available: обновлён (state open→open, comments 5→6).
• FlashML-org/FreeToken:main: bd372b63 → 2757bb5f — feat(server)!: add --gpu to choose the GPU on multi-GPU machines (feat(server)!: add --gpu to choose the GPU on multi-GPU machines FlashML-org/FreeToken#117).
• Upstream main: релевантные изменения bd372b63 → 2757bb5f — files: python/freetoken/engine/engine.py, python/freetoken/kernel/triton/e4m3_compat.py, python/freetoken/kernel/triton/sampling.py, python/freetoken/moe/bench_profile.py, python/freetoken/moe/benchbw.py, python/freetoken/moe/host_banks.py, python/freetoken/scheduler/scheduler.py.

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T03:13:46+00:00).

• FreeToken Roadmap (2026): comments 4→5.
• Новый релевантный issue: GGUF loader only supports gemma4 — Qwen2 / Qwen2.5 GGUF fails with "architecture 'qwen2' is not supported".
• CUDA: XOR swizzle flash attn K,V smem fp16 tiles: обновлён (state open→open, comments 22→23).
• Новый релевантный issue: Eval bug: Excessive compute buffer reservation in MTP draft context on ROCm HIP unnecessarily reduces fitted context size.
• Prefill hangs on RTX 4060 Ti (SM89) for prompts beyond warmup length: обновлён (state open→open, comments 2→3).

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T04:55:02+00:00).

• GGUF: support all quant types, add qwen35moe: новый commit 6f6c8640 → b2f84751.
• Новый релевантный issue: Add explicit detection/support for compressed-tensors NVFP4A16 DeepSeek-V4 REAM checkpoints.
• Новый релевантный issue: Qwen3.5 MoE modelopt_mixed loader assumes shared experts use routed-expert NVFP4 quantization.
• [Bug]: Detect hardware fails with cudaErrorNotSupported on NVIDIA RTX 4060 Lapt: обновлён (state open→closed, comments 2→2).
• GGUF loader only supports gemma4 — Qwen2 / Qwen2.5 GGUF fails with "architecture 'qwen2' is not supported": обновлён (state open→closed, comments 0→1).
• CUDA: extend MOE fusion to specdec, earlier MOE glu fusion and topk-router fusion were restricted to 1 token: обновлён (state open→open, comments 3→4).
• Новый релевантный issue: Eval bug: MTP retains inter-request state causing non-deterministic output and model degradation (Qwen3.6-35B-A3B-MTP).
• Новый релевантный issue: Arbitrary Code Execution in PyTorch via torch.package.PackageImporter.load_pickle.
• Новый релевантный issue: torch.distributions.kl_divergence Fails with MultivariateNormal in Dynamo Due to _infer_size Type Error.
• Новый релевантный issue: Model catalog shows models the local machine cannot run (no capability pre-check before display).

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T07:09:50+00:00).

• Новый релевантный pr: ggml-cuda: enable MMA FlashAttention for head dim 256 on AMD RDNA.
• Новый релевантный pr: CI: fix workflow_dispatch runner fallback for PyTorch wheel _ci builds + tolerate newer ninja.
• [ROCm][inductor] Fix Inductor compiler selection for ROCm on Windows: обновлён (state open→open, comments 5→7).
• Новый релевантный pr: HIP: RDNA 4 MUL_MAT optimizations.
• Новый релевантный pr: CUDA: enable the CUB path on HIP via hipCUB.
• Новый релевантный issue: Feature request: AMD GPU support.
• Новый релевантный issue: Vulkan: model load hangs indefinitely on RDNA4 (gfx1201) with plain Unsloth Q6_K quant, or OOMs host with --direct-io.
• Новый релевантный pr: cuda : dynamic MMVQ nwarps for narrow matrices.

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T09:01:32+00:00).

• GGUF: support all quant types, add qwen35moe: новый commit b2f84751 → dd48aacb.
• Новый релевантный pr: Make the drafter probabilistic and the target verify by rejection sampling for simple draft and MTP.
• Новый релевантный issue: Eval bug: qwen35moe (hybrid recurrent/attention) models produce empty generation when prompt exceeds ~16K tokens via llama-server (Metal).
• Новый релевантный pr: fix(engine): load prefill triton kernels at startup, not mid-request.
• Новый релевантный issue: CUDA OoM with larger --kv-reserve-tokens (workaround: empiric lower --memory-ratio).
• Новый релевантный pr: Vulkan: add hoisting support for row IDs and expert count in shaders.
• Новый релевантный pr: server: add recurrent state shrink/expand for prompt cache (#22746).
• Новый релевантный pr: vulkan : transpose A-matrix data layout for K-quant mul_mat performance.
• Новый релевантный issue: Silent GTT fallback when context + MTP exceeds VRAM — no error at load, massive throughput collapse at runtime.
• Новый релевантный issue: Eval bug: --batch-size and --ubatch-size ignored when set to 1024/1024.
• Новый релевантный issue: draft-mtp speculative decoding not activating on Turing (sm_75) with hybrid SSM+attention model (Qwen3.6-35B-A3B).
• Новый релевантный pr: WIP: kv-cache apply_ubtach change from O(n) to O(1).
• Новый релевантный issue: Eval bug: vulkan performance drop pp512 560 -> 180.
• Новый релевантный pr: llama + spec: MTP Support.
• CUDA: enable the CUB path on HIP via hipCUB: обновлён (state open→open, comments 46→47).

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T10:51:05+00:00).

• HIP/ROCm backend produces corrupted output on gfx1151: two dense architectures fail, Vulkan is correct with byte-identical flags: обновлён (state open→open, comments 7→8).
• Eval bug: Windows HIP gfx1201: Qwen3.6-35B-A3B NVFP4 loads, then first MUL_MAT fails with ROCm invalid argument: обновлён (state open→open, comments 0→4).
• Новый релевантный issue: Long prefill (>48k tokens) crashes the whole server — CUDA driver error: device not ready in fla/l2norm kernel.
• Новый релевантный issue: [Issue] Linux gfx103X: GPU not accessible to runner — rocminfo exits 8 / "No GPU found matching.
• Новый релевантный issue: torch.cholesky_inverse fails with HIPBLAS_STATUS_ALLOC_FAILED at batch 20 on ROCm due to fixed hipBLAS workspace.
• Новый релевантный pr: [ROCm] Enable CK SDPA on gfx1200/gfx1201 (RDNA4).
• Новый релевантный issue: [Issue]: Access violation (0xC0000005) in bmm/baddbmm on gfx1201 (RX 9070 XT) under Windows, crashes in both hipBLASLt and rocBLAS.
• Новый релиз: ggml-org/llama.cpp v0.3.0.

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T11:11:49+00:00).

• Новый явно отслеживаемый external issue: [Issue]: rocBLAS bf16/f16 GEMM fails with hipErrorInvalidValue on 2nd call, gfx1201 / Windows / ROCm 7.14. High relevance.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 8024:3880E4:B12B:25494:6A8D788F and timestamp 2026-08-25 11:12:15 UTC. For…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T13:13:08+00:00).

• feat(rocm): add RDNA3 and RDNA4 runtime foundation: обновлены описание/review/discussion (comments 1→2, review 0→0).
• Новый релевантный issue: Feature: Predictive Expert Offloading.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID F440:3A3DE2:D8F1C6:2D7E43E:6A8D950E and timestamp 2026-08-25 13:13:50 UTC.…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T15:07:11+00:00).

• FreeToken Roadmap (2026): comments 5→6.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 0407:272BC6:1AC4A9F:5964A5E:6A8DAFBB and timestamp 2026-08-25 15:07:39 UTC…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T16:57:34+00:00).

• Model catalog shows models the local machine cannot run (no capability pre-check before display): обновлён (state open→open, comments 0→1).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 283F:2B3BAA:2009A4A:6A360EB:6A8DC9A1 and timestamp 2026-08-25 16:58:09 UTC…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T18:58:44+00:00).

• Новый релевантный pr: fix(deps): use PyTorch 2.13 to fix SM89 hangs.
• Новый релевантный issue: Apodex-1.1-mini-NVFP4 fails: dense_quant hardcoded to NVFP4 when experts are NVFP4, but shared_expert is FP8 (modelopt MIXED_PRECISION).
• Difficult to load model despite more than enough vram/ram available: обновлён (state open→open, comments 6→7).
• Новый релевантный issue: FP8 row-wise _scaled_mm hangs on SM89 with PyTorch 2.11.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID C819:1EE7B3:2EC3E0B:9D340E1:6A8DE60D and timestamp 2026-08-25 18:59:25 UTC…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T20:47:37+00:00).

• fix(deps): use PyTorch 2.13 to fix SM89 hangs: обновлён (state open→open, comments 0→0).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 7830:16ACDE:3C41178:CA29304:6A8DFF87 and timestamp 2026-08-25 20:48:07 UTC…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T22:46:42+00:00).

• GGUF: support all quant types, add qwen35moe: новый commit dd48aacb → 9952a39d.
• Maxritz/FreeToken-rocm-test:main: 45675e47 → 3a5616ec — Merge pull request Tweak globals so that LLVM builds with the proper shared library settings. ROCm/TheRock#1 from Castoff995/codex/gfx1201-offload-decode.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID CC18:2EB04F:4D5504F:100358CF:6A8E1B72 and timestamp 2026-08-25 22:47:14 UT…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-26T01:59:36+00:00).

• GGUF: support all quant types, add qwen35moe: новый commit 9952a39d → 7da6b77e.
• feat(rocm): enable native GGUF kernels on RDNA4: обновлены описание/review/discussion (comments 0→1, review 0→0).
• Requests longer than the KV pool are queued forever with no error — and --moe-cache-auto leaves only ~8k tokens of KV on a 96GB GPU: comments 1→2.
• Новый релевантный pr: feat(kvcache): 8-bit KV cache (q8_0 / fp8_e4m3) behind --kv-cache-dtype.
• FlashML-org/FreeToken:main: 2757bb5f → f7c31e92 — fix(fla): stop l2norm recompiling per token count.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 7431:947DB:5671D36:1210F83B:6A8E48B1 and timestamp 2026-08-26 02:00:17 UTC…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-26T03:19:34+00:00).

• feat(rocm): add RDNA3 and RDNA4 runtime foundation: обновлены описание/review/discussion (comments 2→3, review 0→0).
• feat(rocm): serve on AMD GPUs through the HIP toolchain: новый commit 7e1d7421 → 5eec2582.
• Новый релевантный issue: Qwen3.6-35B-A3B hits KV-cache limit during prolonged reasoning on LeetCode problem.
• Model catalog shows models the local machine cannot run (no capability pre-check before display): обновлён (state open→open, comments 1→2).
• Новый релевантный issue: [AppImage] Global LD_LIBRARY_PATH export breaks system-curl child processes (engine installer fails on non-Ubuntu distros).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID B00B:2DC8AF:65DCB0D:15086008:6A8E5B63 and timestamp 2026-08-26 03:20:03 UT…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-26T04:57:21+00:00).

• Новый релевантный issue: [Linux] AppImage blank window / WebKitWebProcess abort: "Could not create default EGL display: EGL_BAD_PARAMETER" (bundled libwayland-client on Mesa 25+ hosts).
• Новый релевантный issue: Loading nvidia/Qwen3.6-35B-A3B-NVFP4 fails: "Incomplete fp8 fusions: ['model.layers.37.linear_attn.in_proj_qkvz']".
• Qwen3.6-35B-A3B hits KV-cache limit during prolonged reasoning on LeetCode problem: обновлён (state open→open, comments 0→0).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID D828:1EDC:4C5ACCD:FD5C08A:6A8E724C and timestamp 2026-08-26 04:57:48 UTC. …

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-26T07:10:22+00:00).

• GGUF: support all quant types, add qwen35moe: новый commit 7da6b77e → bb432e8c.
• FreeToken Roadmap (2026): body edited.
• Новый релевантный pr: qwen3_5_moe: load mixed-precision compressed-tensors checkpoints (unsloth NVFP4+FP8).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID CAC4:33DF9:7ABF9D8:19628807:6A8E9187 and timestamp 2026-08-26 07:11:03 UTC…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-26T09:04:06+00:00).

• Bug in PyTorch 2.11 causes SM89 hangs: обновлён (state open→open, comments 0→0).
• feat(models): support TP for qwen3_5_moe: обновлён (state open→open, comments 2→3).
• FlashML-org/FreeToken:main: f7c31e92 → 9ef36513 — chore(assets): update wechat group QR code.
• Maxritz/FreeToken-rocm-test:main: 3a5616ec → 89f14790 — Merge pull request Enable the GH build. ROCm/TheRock#2 from PialGhosh2233/main.
• PialGhosh2233/FreeToken-rocm-gfx1200:gfx1200-moe-gguf: 544b83df → a338f49b — docs: Gemma-4-26B-A4B QAT q4_0 GGUF results on gfx1200.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 0846:30A2EB:30974E:A35D01:6A8EAC23 and timestamp 2026-08-26 09:04:35 UTC. …

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-26T10:52:44+00:00).

• Новый релевантный pr: feat(models): add Qwen3-Next-80B-A3B support.
• feat(models): support TP for qwen3_5_moe: обновлён (state open→open, comments 3→5).
• Новый релиз: ROCm/TheRock therock-10.0.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID B008:2BB36A:73A408:77B54F:6A8EC59E and timestamp 2026-08-26 10:53:18 UTC. …

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-26T13:18:13+00:00).

• FreeToken Roadmap (2026): comments 6→7.
• Feature request: AMD GPU support: обновлён (state open→open, comments 11→12).
• feat(models): support TP for qwen3_5_moe: обновлён (state open→open, comments 5→7).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 0400:15153F:A5D123:230714B:6A8EE7C0 and timestamp 2026-08-26 13:18:56 UTC.…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-26T15:55:00+00:00).

• [Linux] AppImage blank window / WebKitWebProcess abort: "Could not create default EGL display: EGL_BAD_PARAMETER" (bundled libwayland-client on Mesa 25+ hosts): обновлён (state open→open, comments 0→1).
• feat(models): support TP for qwen3_5_moe: обновлён (state open→open, comments 7→8).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID D82D:361ECC:232D6C:2C6E23:6A8F0C76 and timestamp 2026-08-26 15:55:34 UTC. …

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-09-29T22:17:53+00:00).

• Новый релевантный pr: fix(windows): build the extensions and JIT kernels with MSVC.
• Новый релевантный pr: feat(engine): --vram-reserve-mb keeps VRAM free at the largest prefill.
• Новый релевантный pr: feat(rotary): size the rope table for YaRN and add --hf-overrides.
• Новый релевантный issue: QSA block top-k runs a 1M-token window on one program per row.
• Новый релевантный pr: perf(qsa): reduce block top-k candidates in levels past 16 splits.
• Новый релевантный issue: Serving Qwen3.8-Flash-Next's 1M-token window on one 24 GB card.
• Новый релевантный issue: feat(server): add a server-side thinking override.
• Build on Windows?: обновлён (state open→open, comments 2→3).
• Новый релевантный issue: On Windows the engine sizes its pools as if no other process held any of the card.
• Новый релевантный issue: A hybrid model's cached prefix cannot be resumed far back: no GDN snapshot survives below a prompt's last chunk.
• Новый релевантный pr: fix(server): keep a silent stream alive and notice a client that left.
• Новый релевантный issue: A stream that is silent through a long prefill is dropped by Node clients and prefilled for nobody.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 9433:48608:A5371C:223DC4E:6ABC392C and timestamp 2026-09-29 22:18:20 UTC. …

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-09-30T05:35:41+00:00).

• QSA block top-k runs a 1M-token window on one program per row: обновлён (state open→closed, comments 0→0).
• perf(qsa): reduce block top-k candidates in levels past 16 splits: обновлён (state open→closed, comments 0→1).
• FlashML-org/FreeToken:main: 0d652e73 → eba7120e — feat(install): record the install root as FREETOKEN_HOME in environment.d (feat(install): record the install root as FREETOKEN_HOME in environment.d FlashML-org/FreeToken#576).
• Upstream main: релевантные изменения 0d652e73 → eba7120e — files: python/freetoken/kernel/triton/qsa/topk.py.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID B84D:1EE698:DC9A3:2E331D:6ABC9FD8 and timestamp 2026-09-30 05:36:24 UTC. F…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-09-30T12:38:14+00:00).

• Новый релевантный issue: RTX 2050 4GB: Qwen3.6-35B-A3B-NVFP4 runs at ~21 tok/s with single-layer CPU MoE prefill buffer.
• Новый релевантный pr: feat(kvcache): keep prefill-chunk GDN checkpoints in a pinned host bank.
• Bug: AssertionError: Not enough memory for KV cache when serving Qwen3.8-27B-NVFP4 on Windows: обновлён (state open→open, comments 0→1).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 1413:1B16ED:32F76E:A90AD2:6ABD02DB and timestamp 2026-09-30 12:38:51 UTC. …

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-09-30T19:22:58+00:00).

• Новый релевантный issue: Feature: Predictive Expert Offloading.
• Новый релевантный pr: perf(linear): run a single-row bf16 projection as one GEMV.
• Новый релевантный issue: The computer on the local area network cannot be accessed.
• Новый релевантный pr: fix(windows): serve over loopback ZMQ and a selector event loop.
• fix(windows): build the extensions and JIT kernels with MSVC: обновлён (state open→open, comments 0→0).
• perf(qsa): reduce block top-k candidates in levels past 16 splits: обновлён (state closed→closed, comments 1→1).
• Новый релевантный pr: fix(engine): size memory pools against the VRAM other processes leave free on Windows.
• Новый релевантный pr: fix(windows): read checkpoints without O_DIRECT, posix_fadvise or PROT_READ.
• Qwen3.8-Flash-Next-NVFP4 cannot start on a 12 GB RTX 3060: ~9.9 GiB of unquantized (BF16) attention/embedding weights leave no room for MoE cache + KV: обновлён (state open→open, comments 4→5).
• RTX 3090 device loss (Xid 79) at --max-prefill-length 8192 in offload mode — known limitation, or host-specific?: обновлён (state open→open, comments 2→3).
• feat(kvcache): keep prefill-chunk GDN checkpoints in a pinned host bank: обновлён (state open→open, comments 0→1).
• Новый релевантный issue: An offloaded prefill streams every expert of each layer, however short the chunk.
• Новый релевантный pr: perf(moe): dequant nvfp4 prefill experts by bit placement and load activations contiguously.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 0841:2CC2EC:67024:723C1:6ABD61B5 and timestamp 2026-09-30 19:23:33 UTC. Fo…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-09-30T23:52:30+00:00).

• feat(rotary): size the rope table for YaRN and add --hf-overrides: обновлён (state open→open, comments 0→1).
• Feature: Predictive Expert Offloading: обновлён (state open→open, comments 2→3).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID DC21:1F03:783509:18CA8FE:6ABDA0E2 and timestamp 2026-09-30 23:53:06 UTC. F…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-01T05:59:01+00:00).

• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 280E:0C59:10BD24:181AE3:6ABDF6C6 and timestamp 2026-10-01 05:59:34 UTC. Fo…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-01T13:20:11+00:00).

• Feature: Predictive Expert Offloading: обновлён (state open→open, comments 3→4).
• perf(linear): run a single-row bf16 projection as one GEMV: обновлён (state open→open, comments 0→0).
• fix(windows): read checkpoints without O_DIRECT, posix_fadvise or PROT_READ: обновлён (state open→open, comments 0→0).
• fix(windows): serve over loopback ZMQ and a selector event loop: обновлён (state open→open, comments 0→0).
• fix(windows): build the extensions and JIT kernels with MSVC: обновлён (state open→open, comments 0→0).
• feat(engine): --vram-reserve-mb keeps VRAM free at the largest prefill: обновлён (state open→open, comments 0→0).
• feat(rotary): size the rope table for YaRN and add --hf-overrides: обновлён (state open→closed, comments 1→3).
• fix(server): keep a silent stream alive and notice a client that left: обновлён (state open→open, comments 0→0).
• FlashML-org/FreeToken:main: eba7120e → 235e2016 — feat(rotary): size the rope table for YaRN and add --hf-overrides (feat(rotary): size the rope table for YaRN and add --hf-overrides FlashML-org/FreeToken#573).
• Upstream main: релевантные изменения eba7120e → 235e2016 — files: python/freetoken/engine/config.py, python/freetoken/engine/engine.py.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 8C2C:15094C:6FE530:16E95F6:6ABE5E29 and timestamp 2026-10-01 13:20:41 UTC.…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-01T19:33:14+00:00).

• Новый релевантный pr: fix(cuda): support Turing GPUs.
• Новый релевантный pr: feat(rocm): run on RDNA3/RDNA4 with torch 2.14/2.15 and ROCm 7.14/10.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 841B:2FD95D:8690D0:1B9F020:6ABEB5A0 and timestamp 2026-10-01 19:33:52 UTC.…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-01T23:54:12+00:00).

• Новый релевантный pr: fix(ple): fence disk staging against in-flight lookups.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID B7C9:2407CC:1CB4D4:262F99:6ABEF2BD and timestamp 2026-10-01 23:54:37 UTC. …

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-02T05:41:20+00:00).

• fix(ple): fence disk staging against in-flight lookups: обновлён (state open→closed, comments 0→0).
• FlashML-org/FreeToken:main: 235e2016 → d3512b43 — fix(ple): fence disk staging against in-flight lookups (fix(ple): fence disk staging against in-flight lookups FlashML-org/FreeToken#588).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID BC82:3804D2:237606:76230C:6ABF4419 and timestamp 2026-10-02 05:41:45 UTC. …

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-02T12:39:16+00:00).

• Новый релевантный pr: fix(checkpoint): resolve HF repo Ids before conversion.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 3805:2B00BF:D3AB8:103004:6ABFA614 and timestamp 2026-10-02 12:39:48 UTC. F…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-02T19:21:01+00:00).

• fix(windows): serve over loopback ZMQ and a selector event loop: обновлён (state open→open, comments 0→0).
• Bug: AssertionError: Not enough memory for KV cache when serving Qwen3.8-27B-NVFP4 on Windows: обновлён (state open→open, comments 1→2).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 4818:F2D31:111865F:37A411D:6AC0044E and timestamp 2026-10-02 19:21:50 UTC.…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-02T23:48:42+00:00).

• feat(kvcache): keep prefill-chunk GDN checkpoints in a pinned host bank: обновлён (state open→open, comments 1→2).
• Новый релевантный pr: perf(model): fuse hc combine with the next block's RMSNorm.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 081C:366C2D:D93CA:2D76E0:6AC04301 and timestamp 2026-10-02 23:49:21 UTC. F…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-03T05:21:21+00:00).

• feat(kvcache): keep prefill-chunk GDN checkpoints in a pinned host bank: обновлён (state open→open, comments 2→3).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 9C31:2A54B5:1DCC3:63ABB:6AC090E8 and timestamp 2026-10-03 05:21:44 UTC. Fo…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-03T11:42:13+00:00).

• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 7807:2638C8:36CEA5:B589E6:6AC0EA2E and timestamp 2026-10-03 11:42:38 UTC. …

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-03T16:19:33+00:00).

• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 6841:198B6C:152CB3C:45C13DA:6AC12B3A and timestamp 2026-10-03 16:20:10 UTC…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-03T21:19:03+00:00).

• Новый релевантный pr: feat(server): add ft info and a pre-load memory preflight.
• Новый релевантный pr: feat(qwen4_exp): support ModelOpt QAD checkpoints (NVFP4 PLE tables, MXFP8 dense).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 401F:1941DC:DA80CA:2D4C801:6AC17166 and timestamp 2026-10-03 21:19:35 UTC.…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-04T00:39:58+00:00).

• feat(server): add ft info and a pre-load memory preflight: обновлён (state open→open, comments 0→1).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID E3C7:2F195D:438424:DD3723:6AC1A076 and timestamp 2026-10-04 00:40:22 UTC. …

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-04T08:28:33+00:00).

• Новый релевантный issue: [Windows][Desktop] WeightLoadError: invalid python storage with Qwen3.6 NVFP4 and gpt-oss-20b MXFP4.
• Новый релевантный pr: feat(engine): --embed-device cpu keeps the input embedding in RAM.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 1448:19B002:1C7904C:5EC8E76:6AC20E50 and timestamp 2026-10-04 08:29:04 UTC…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-04T15:23:38+00:00).

• Новый релевантный pr: fix(engine): reserve the pool's dummy page in the --moe-cache-auto KV floor.
• Новый релевантный pr: fix(kernels): drop the D2H sync from the varlen GDN/KDA prefill conv.
• Новый релевантный pr: perf(ple): fuse the n-gram row-id hash into one Triton kernel.
• Новый релевантный pr: feat(moe): serve NVFP4 and MXFP4 experts fused through the expert banks.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 4A8A:8DB83:263679F:7F49B01:6AC26FA1 and timestamp 2026-10-04 15:24:17 UTC.…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-04T19:39:36+00:00).

• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID C009:8DB83:354BA3E:B161B1D:6AC2AB90 and timestamp 2026-10-04 19:40:00 UTC.…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-04T23:11:57+00:00).

• feat(moe): serve NVFP4 and MXFP4 experts fused through the expert banks: обновлён (state open→closed, comments 0→0).
• FlashML-org/FreeToken:main: d3512b43 → 189c8ad0 — feat(moe): serve NVFP4 and MXFP4 experts fused through the expert banks (feat(moe): serve NVFP4 and MXFP4 experts fused through the expert banks FlashML-org/FreeToken#601).
• Upstream main: релевантные изменения d3512b43 → 189c8ad0 — commits: feat(moe): serve NVFP4 and MXFP4 experts fused through the expert banks (feat(moe): serve NVFP4 and MXFP4 experts fused through the expert banks FlashML-org/FreeToken#601) | files: python/freetoken/engine/engine.py, python/freetoken/models/qwen3_5_moe/weight.py, python/freetoken/moe/expert_banks.py, python/freetoken/moe/offload_cache.py.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID F812:2CA5B0:5EB16:734A7:6AC2DD5A and timestamp 2026-10-04 23:12:26 UTC. Fo…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-05T05:44:54+00:00).

• FreeToken Roadmap (2026): body edited.
• Новый релевантный pr: fix(cpu-moe): honor padded fp8 scale strides, add avx512f tier and float64 parity (builds on #36).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID C01A:DF67F:4CDB124:FE53A4A:6AC33977 and timestamp 2026-10-05 05:45:27 UTC.…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-05T14:44:50+00:00).

• Новый релевантный pr: feat(models): support deepseek v4.1 with native fp8-fp4 kv storage.
• Новый релевантный issue: [Model] DeepSeek-V4.1-Flash (deepseek_v41) — hybrid CPU-MoE makes it viable on a single GPU, and Engram is already isolated in 2 shards.
• FP8 block-quantized MoE models have no usable kernel under moe_strategy=offload (0.1.3): обновлён (state open→open, comments 3→4).
• fix(cpu-moe): honor padded fp8 scale strides, add avx512f tier and float64 parity (builds on #36): обновлён (state open→open, comments 1→2).
• [Desktop] weights failed to load: Qwen3.6-27B NVFP4: обновлён (state open→closed, comments 0→1).
• [Windows][Desktop] WeightLoadError: invalid python storage with Qwen3.6 NVFP4 and gpt-oss-20b MXFP4: обновлён (state open→open, comments 0→1).
• FlashML-org/FreeToken:main: 189c8ad0 → 581fdca3 — fix(server): report the backend kv page size in stats (fix(server): report the backend kv page size in stats FlashML-org/FreeToken#599).
• Новый релиз: ROCm/TheRock therock-10.1.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 8C45:124843:3F117F:D4444F:6AC3B7FE and timestamp 2026-10-05 14:45:18 UTC. …

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-05T22:39:39+00:00).

• Новый релевантный pr: fix(server): keep Anthropic inline system messages in place for prefix reuse.
• Новый релевантный issue: Windows wheels resolve unpinned torch dependency to CPU-only build: 'cannot use CUDA device 0: only 0 device(s) visible'.
• Новый релевантный issue: DeepSeek-V4-Flash-0731-NVFP4 (nvidia) fails on 0.1.3+gfa0e9537d: WeightLoadError KeyError 'layers.0.ffn.experts.0.w1.scale'.
• [Windows][Desktop] WeightLoadError: invalid python storage with Qwen3.6 NVFP4 and gpt-oss-20b MXFP4: обновлён (state open→open, comments 1→2).
• Новый релевантный pr: perf(pre-sm70): pipeline the block-fp8 GEMMs and halve the extend-attention K/V re-reads on Pascal.
• FlashML-org/FreeToken:main: 581fdca3 → 555efd89 — fix(kvcache): clone inserted radix keys to release request buffers (fix(kvcache): clone inserted radix keys to release request buffers FlashML-org/FreeToken#559).
• Новый релиз: ggml-org/llama.cpp v0.6.0.
• Upstream main: релевантные изменения 581fdca3 → 555efd89 — files: python/freetoken/kvcache/hybrid_radix_cache.py, python/freetoken/kvcache/radix_cache.py, python/freetoken/kvcache/swa_radix_cache.py.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 67C0:3CC60:553DEF:61697C:6AC4274A and timestamp 2026-10-05 22:40:10 UTC. F…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-06T06:24:26+00:00).

• Новый релевантный pr: perf(moe): Improved fast_index_copy_multi_jit for NVLink-C2C hosts (GH200: 222 → 412 GB/s).
• DeepSeek-V4-Flash-0731-NVFP4 (nvidia) fails on 0.1.3+gfa0e9537d: WeightLoadError KeyError 'layers.0.ffn.experts.0.w1.scale': обновлён (state open→open, comments 0→1).
• Новый релевантный pr: fix(models): preserve glm-5.2 router correction bias in fp32.
• Windows wheels resolve unpinned torch dependency to CPU-only build: 'cannot use CUDA device 0: only 0 device(s) visible': обновлён (state open→open, comments 0→1).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 3986:385E0D:2365920:700D4CA:6AC49442 and timestamp 2026-10-06 06:25:06 UTC…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-06T15:13:59+00:00).

• fix(rocm): route tensor parallel communication through RCCL: обновлены описание/review/discussion (comments 4→5, review 0→0).
• Unhandled CUDA OOM in prefill expert workspace kills the engine; server keeps accepting requests that never return (root cause behind #20-style headless state; likely also #72): comments 2→3.
• Новый релевантный issue: fix(server): allow overriding the distributed port.
• Новый релевантный pr: fix(server)!: answer /health with 503 until the server accepts requests.
• fix(engine): size memory pools against the VRAM other processes leave free on Windows: обновлён (state open→open, comments 0→0).
• fix(windows): serve over loopback ZMQ and a selector event loop: обновлён (state open→open, comments 0→0).
• perf(moe): dequant nvfp4 prefill experts by bit placement and load activations contiguously: обновлён (state open→open, comments 0→0).
• fix(windows): build the extensions and JIT kernels with MSVC: обновлён (state open→open, comments 0→0).
• feat(engine): --vram-reserve-mb keeps VRAM free at the largest prefill: обновлён (state open→open, comments 0→0).
• Новый релевантный pr: fix(moe): route moe_align to the atomic-free staged path below sm_70.
• Новый релевантный issue: GLM-5.3-Flash-NVFP4 CUDA OOM during load on RTX 5070 Ti 16 GB (sm_120).
• Qwen3.8-Flash-Next-NVFP4 cannot start on a 12 GB RTX 3060: ~9.9 GiB of unquantized (BF16) attention/embedding weights leave no room for MoE cache + KV: обновлён (state open→open, comments 5→6).
• Новый релевантный issue: [Bug] b12x NVFP4 backend on sm_120 (RTX 5090) serves batch 1 only: not CUDA-graph capturable, and the worker dies with a tile-config ValueError at batch >= 2 (Qwen3.8-Flash-Next-NVFP4).
• perf(moe): Improved fast_index_copy_multi_jit for NVLink-C2C hosts (GH200: 222 → 412 GB/s): обновлён (state open→open, comments 1→2).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 6C13:18C7C3:47CBA2:E9F219:6AC5105E and timestamp 2026-10-06 15:14:38 UTC. …

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-06T21:00:05+00:00).

• FreeToken Roadmap (2026): body edited.
• Новый релевантный pr: feat(attention): hand-written GQA-fused extend kernel for pre-sm70 (1.8x).
• Новый релевантный pr: perf(fla): fix the pre-sm70 GDN chunk-kernel launch configs (7.4x / 11.3x).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 43C4:DEF79:8DD1D:1D23B9:6AC5616F and timestamp 2026-10-06 21:00:31 UTC. Fo…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-07T01:34:15+00:00).

• perf(moe): Improved fast_index_copy_multi_jit for NVLink-C2C hosts (GH200: 222 → 412 GB/s): обновлён (state open→open, comments 2→3).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 781C:36FE9D:3B7DE9:C45B31:6AC5A1C0 and timestamp 2026-10-07 01:34:56 UTC. …

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-07T08:42:27+00:00).

• Новый релевантный pr: perf(moe): spin-wait the cpu moe pool between layers.
• perf(moe): Improved fast_index_copy_multi_jit for NVLink-C2C hosts (GH200: 222 → 412 GB/s): обновлён (state open→closed, comments 3→5).
• Новый релевантный pr: fix(models): better support for mixed-precision compressed-tensors NVFP4.
• FlashML-org/FreeToken:main: 555efd89 → 127c8566 — feat(deepseek-v41): support text and vision serving (feat(deepseek-v41): support text and vision serving FlashML-org/FreeToken#603).
• Upstream main: релевантные изменения 555efd89 → 127c8566 — commits: perf(moe): Improved fast_index_copy_multi_jit for NVLink-C2C hosts (perf(moe): Improved fast_index_copy_multi_jit for NVLink-C2C hosts (GH200: 222 → 412 GB/s) FlashML-org/FreeToken#610); perf(moe): spin-wait the cpu moe pool between layers (perf(moe): spin-wait the cpu moe pool between layers FlashML-org/FreeToken#619) | files: python/freetoken/engine/cache_budget.py, python/freetoken/engine/config.py, python/freetoken/engine/engine.py, python/freetoken/kernel/fast_index_copy.py, python/freetoken/kernel/triton/dsv4/fp8_linear.py, python/freetoken/kernel/triton/dsv4/fused_moe.py, python/freetoken/kernel/triton/dsv4/router.py, python/freetoken/kernel/triton/dsv41/init.py.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 4845:1683C0:FBE78:16CDA3:6AC60612 and timestamp 2026-10-07 08:42:58 UTC. F…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.