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":"2026-10-14T03:30:34Z","message":"perf(moe): split hybrid decode routes inside the ensure kernel (FlashML-org#685)","repo":"FlashML-org/FreeToken","sha":"f02fbb16a213f5c81094e4017c988d41f0164141","url":"FlashML-org@f02fbb16a213f5c81094e4017c988d41f0164141"},"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#139%22:%7B%22body_hash%22:%22b1f1d04141276499%22,%22comments%22:2,%22kind%22:%22issue%22,%22number%22:139,%22repo%22:%22FlashML-org/FreeToken%22,%22state%22:%22open%22,%22title%22:%22%60Qwen3.8-27B-NVFP4%60 fails to load — RuntimeError: CUDA driver error: device not ready","updated_at":"2026-10-09T15:47:36Z","url":"https://github.com/FlashML-org/FreeToken/issues/139"},"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":"2026-10-13T14:27:17Z","url":"https://github.com/FlashML-org/FreeToken/issues/301"},"FlashML-org/FreeToken#378":{"body_hash":"fccfe0cb568d91d9","comments":2,"kind":"pr","number":378,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(rocm): make CPU/Hybrid MoE graph replay safe","updated_at":"2026-10-13T03:02:18Z","url":"https://github.com/FlashML-org/FreeToken/pull/378"},"FlashML-org/FreeToken#399":{"body_hash":"227a0901fd907e75","comments":4,"kind":"pr","number":399,"repo":"FlashML-org/FreeToken","state":"closed","title":"fix(cpu-moe): honor padded fp8 scale strides, add avx512f tier and float64 parity (builds on FlashML-org#36)","updated_at":"2026-10-12T05:18:45Z","url":"https://github.com/FlashML-org/FreeToken/pull/399"},"FlashML-org/FreeToken#408":{"body_hash":"c59b1167f4e8a6b6","comments":2,"kind":"pr","number":408,"repo":"FlashML-org/FreeToken","state":"open","title":"feat(kvcache): add nvfp4 kv quantization","updated_at":"2026-10-13T15:26:36Z","url":"https://github.com/FlashML-org/FreeToken/pull/408"},"FlashML-org/FreeToken#460":{"body_hash":"efe5429c0d3bc8aa","comments":3,"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":"2026-10-09T06:26:58Z","url":"https://github.com/FlashML-org/FreeToken/pull/460"},"FlashML-org/FreeToken#490":{"body_hash":"e7789dd22c580c87","comments":1,"kind":"issue","number":490,"repo":"FlashML-org/FreeToken","state":"open","title":"[Bug] Corrupted Triton JIT cache crash-loops the engine: worker dies on kernel load, auto-restart can never heal disk poison","updated_at":"2026-10-11T05:36:36Z","url":"https://github.com/FlashML-org/FreeToken/issues/490"},"FlashML-org/FreeToken#491":{"body_hash":"0d2c659b585fce9d","comments":1,"kind":"pr","number":491,"repo":"FlashML-org/FreeToken","state":"open","title":"refactor(moe): extract hybrid decode orchestration","updated_at":"2026-10-09T04:25:05Z","url":"https://github.com/FlashML-org/FreeToken/pull/491"},"FlashML-org/FreeToken#534":{"body_hash":"265c914cd1ae1994","comments":5,"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":"2026-10-12T05:19:36Z","url":"https://github.com/FlashML-org/FreeToken/issues/534"},"FlashML-org/FreeToken#541":{"body_hash":"4de9324ef8bd9366","comments":0,"kind":"issue","number":541,"repo":"FlashML-org/FreeToken","state":"open","title":"AMD Roadmap 2026 Q4-2027 Q1","updated_at":"2026-10-12T08:19:15Z","url":"https://github.com/FlashML-org/FreeToken/issues/541"},"FlashML-org/FreeToken#572":{"body_hash":"57a9a477a68f65dc","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":"2026-10-13T14:06:13Z","url":"https://github.com/FlashML-org/FreeToken/pull/572"},"FlashML-org/FreeToken#574":{"body_hash":"b1d2d29110bf684b","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":"2026-10-13T14:06:16Z","url":"https://github.com/FlashML-org/FreeToken/pull/574"},"FlashML-org/FreeToken#575":{"body_hash":"adb2ebcfc13459f5","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":"2026-10-13T14:06:18Z","url":"https://github.com/FlashML-org/FreeToken/pull/575"},"FlashML-org/FreeToken#578":{"body_hash":"994f621bf1d7dcbf","comments":2,"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":"2026-10-13T14:06:20Z","url":"https://github.com/FlashML-org/FreeToken/pull/578"},"FlashML-org/FreeToken#580":{"body_hash":"900b3387f7f0c26a","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":"2026-10-13T14:06:23Z","url":"https://github.com/FlashML-org/FreeToken/pull/580"},"FlashML-org/FreeToken#581":{"body_hash":"70c898e2d004662e","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":"2026-10-13T14:06:27Z","url":"https://github.com/FlashML-org/FreeToken/pull/581"},"FlashML-org/FreeToken#583":{"body_hash":"02f167abe3c51d50","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":"2026-10-13T14:06:32Z","url":"https://github.com/FlashML-org/FreeToken/pull/583"},"FlashML-org/FreeToken#585":{"body_hash":"cb6f48be18636151","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":"2026-10-13T14:06:34Z","url":"https://github.com/FlashML-org/FreeToken/pull/585"},"FlashML-org/FreeToken#613":{"body_hash":"4ab10d67575018ea","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":"2026-10-13T14:06:39Z","url":"https://github.com/FlashML-org/FreeToken/pull/613"},"FlashML-org/FreeToken#621":{"body_hash":"bb378400b3517987","comments":1,"kind":"issue","number":621,"repo":"FlashML-org/FreeToken","state":"open","title":"QSA host KV tier: give each layer its own slot map (decode at 231K 6.8 -> 15.6 tok/s)","updated_at":"2026-10-13T19:36:37Z","url":"https://github.com/FlashML-org/FreeToken/issues/621"},"FlashML-org/FreeToken#624":{"body_hash":"a95d0004d3c436d1","comments":1,"kind":"issue","number":624,"repo":"FlashML-org/FreeToken","state":"open","title":"Load-time MXFP8 requantization of bf16 dense, hyper-connection and lm_head weights","updated_at":"2026-10-13T19:36:42Z","url":"https://github.com/FlashML-org/FreeToken/issues/624"},"FlashML-org/FreeToken#625":{"body_hash":"708e3791f0667376","comments":1,"kind":"issue","number":625,"repo":"FlashML-org/FreeToken","state":"open","title":"CPU MoE: AVX2 int8-activation NVFP4 kernel for CPUs without VNNI","updated_at":"2026-10-13T19:36:44Z","url":"https://github.com/FlashML-org/FreeToken/issues/625"},"FlashML-org/FreeToken#628":{"body_hash":"304197bb3b3c0786","comments":1,"kind":"issue","number":628,"repo":"FlashML-org/FreeToken","state":"open","title":"Pre-sm70: remaining launch-config fixes on the Qwen3.8-Flash-Next path","updated_at":"2026-10-13T19:36:49Z","url":"https://github.com/FlashML-org/FreeToken/issues/628"},"FlashML-org/FreeToken#629":{"body_hash":"c6a8d6e8e144fc4d","comments":1,"kind":"issue","number":629,"repo":"FlashML-org/FreeToken","state":"closed","title":"Upgrade to torch 2.14.1, Triton 3.8 and sglang-kernel 0.4.9","updated_at":"2026-10-13T09:42:10Z","url":"https://github.com/FlashML-org/FreeToken/issues/629"},"FlashML-org/FreeToken#630":{"body_hash":"dc45fbfb34f87f56","comments":0,"kind":"pr","number":630,"repo":"FlashML-org/FreeToken","state":"closed","title":"build(deps): move to torch 2.14.1, triton 3.8 and sglang-kernel 0.4.9","updated_at":"2026-10-10T20:30:22Z","url":"https://github.com/FlashML-org/FreeToken/pull/630"},"FlashML-org/FreeToken#638":{"body_hash":"736c57ac65949207","comments":2,"kind":"pr","number":638,"repo":"FlashML-org/FreeToken","state":"closed","title":"fix(kernel): correct jit tensor matcher checks and pdl arch guards","updated_at":"2026-10-08T20:31:44Z","url":"https://github.com/FlashML-org/FreeToken/pull/638"},"FlashML-org/FreeToken#642":{"body_hash":"c9a6cc0ea2dd8fec","comments":1,"kind":"issue","number":642,"repo":"FlashML-org/FreeToken","state":"open","title":"--moe-cache-auto asserts "cache budget too small" instead of falling back to prefill overlap off","updated_at":"2026-10-13T19:36:50Z","url":"https://github.com/FlashML-org/FreeToken/issues/642"},"FlashML-org/FreeToken#643":{"body_hash":"cbd316198467f1d9","comments":2,"kind":"issue","number":643,"repo":"FlashML-org/FreeToken","state":"open","title":"Live cache rebuild grows MoE before shrinking KV/mamba, so a resize that passes the fit check can OOM","updated_at":"2026-10-13T19:36:52Z","url":"https://github.com/FlashML-org/FreeToken/issues/643"},"FlashML-org/FreeToken#646":{"body_hash":"5237e39827c0b799","comments":1,"kind":"issue","number":646,"repo":"FlashML-org/FreeToken","state":"open","title":"/v1/chat/completions returns 400 for the OpenAI developer role","updated_at":"2026-10-13T19:36:57Z","url":"https://github.com/FlashML-org/FreeToken/issues/646"},"FlashML-org/FreeToken#648":{"body_hash":"38248afb18e09800","comments":1,"kind":"issue","number":648,"repo":"FlashML-org/FreeToken","state":"open","title":"Qwen3.8-Flash-Next: transformers 5.16 drops \p{M} from the pre-tokenizer, so combining-mark text gets different ids","updated_at":"2026-10-13T19:36:59Z","url":"https://github.com/FlashML-org/FreeToken/issues/648"},"FlashML-org/FreeToken#650":{"body_hash":"db0b6755d27ca816","comments":0,"kind":"pr","number":650,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(engine): apply shrinking cache resizes before growing ones in live rebuild (Closes FlashML-org#643)","updated_at":"2026-10-08T19:39:41Z","url":"https://github.com/FlashML-org/FreeToken/pull/650"},"FlashML-org/FreeToken#652":{"body_hash":"4d15e6990632bb46","comments":0,"kind":"pr","number":652,"repo":"FlashML-org/FreeToken","state":"closed","title":"test(kernel): cover Triton activation and norm fallbacks","updated_at":"2026-10-09T06:51:56Z","url":"https://github.com/FlashML-org/FreeToken/pull/652"},"FlashML-org/FreeToken#653":{"body_hash":"2c6f0d56c4b175c9","comments":0,"kind":"pr","number":653,"repo":"FlashML-org/FreeToken","state":"closed","title":"perf(kernel): avoid the activation layout conversion in fp8 moe decode and read mxfp4 gemv weights as column pairs","updated_at":"2026-10-09T07:15:08Z","url":"https://github.com/FlashML-org/FreeToken/pull/653"},"FlashML-org/FreeToken#654":{"body_hash":"4bb77216d9c00104","comments":0,"kind":"pr","number":654,"repo":"FlashML-org/FreeToken","state":"closed","title":"fix(kernel): avoid nan in triton sampling softmax when a chunk starts with -inf","updated_at":"2026-10-10T08:17:57Z","url":"https://github.com/FlashML-org/FreeToken/pull/654"},"FlashML-org/FreeToken#656":{"body_hash":"adcb6a0b948aedd1","comments":0,"kind":"pr","number":656,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(moe): refuse e8m0 scale code 0xFF when MXFP4 expert pieces are read","updated_at":"2026-10-10T01:30:59Z","url":"https://github.com/FlashML-org/FreeToken/pull/656"},"FlashML-org/FreeToken#659":{"body_hash":"88694f7ec72669bd","comments":0,"kind":"pr","number":659,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(kernel): enable disk row stores on windows","updated_at":"2026-10-11T07:24:13Z","url":"https://github.com/FlashML-org/FreeToken/pull/659"},"FlashML-org/FreeToken#660":{"body_hash":"dfe87dbc9da8f9fa","comments":0,"kind":"pr","number":660,"repo":"FlashML-org/FreeToken","state":"open","title":"feat(moe): support gsq q2_0 routed experts","updated_at":"2026-10-11T09:19:13Z","url":"https://github.com/FlashML-org/FreeToken/pull/660"},"FlashML-org/FreeToken#661":{"body_hash":"3673dc784be5d9b3","comments":0,"kind":"issue","number":661,"repo":"FlashML-org/FreeToken","state":"open","title":"Requests stall and time out above --max-running-requests 8; GPU stays at ~28% of TDP and only one CPU thread is busy","updated_at":"2026-10-11T12:50:49Z","url":"https://github.com/FlashML-org/FreeToken/issues/661"},"FlashML-org/FreeToken#662":{"body_hash":"13fb6af2a9f425fd","comments":0,"kind":"pr","number":662,"repo":"FlashML-org/FreeToken","state":"closed","title":"feat(cpu-moe): serve block-fp8 experts on the cpu and hybrid backends","updated_at":"2026-10-12T05:12:42Z","url":"https://github.com/FlashML-org/FreeToken/pull/662"},"FlashML-org/FreeToken#663":{"body_hash":"6d7e00f9f32f8d9c","comments":0,"kind":"issue","number":663,"repo":"FlashML-org/FreeToken","state":"open","title":"sm75 (Turing) fork: 35B MoE on an 8 GB RTX 2080 — 350/25 t/s, 64K context","updated_at":"2026-10-12T06:25:14Z","url":"https://github.com/FlashML-org/FreeToken/issues/663"},"FlashML-org/FreeToken#664":{"body_hash":"ad3c61da62aa944a","comments":0,"kind":"issue","number":664,"repo":"FlashML-org/FreeToken","state":"open","title":"bench bw verdict misses in the other direction: auto stays on offload (ratio ~1.9), while hybrid decodes 21% / 55% faster (Qwen3.8-Flash-Next NVFP4, RTX 4090 + dual Xeon)","updated_at":"2026-10-12T13:13:17Z","url":"https://github.com/FlashML-org/FreeToken/issues/664"},"FlashML-org/FreeToken#667":{"body_hash":"e3b0c44298fc1c14","comments":0,"kind":"pr","number":667,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(rocm): verify native cpu moe graph handshake on rocm 10","updated_at":"2026-10-13T08:50:00Z","url":"https://github.com/FlashML-org/FreeToken/pull/667"},"FlashML-org/FreeToken#669":{"body_hash":"b018acbd1ca7e857","comments":0,"kind":"pr","number":669,"repo":"FlashML-org/FreeToken","state":"closed","title":"build(deps): move to flashlib 0.4.0 and drop the numpy<2.5 cap","updated_at":"2026-10-13T08:24:14Z","url":"https://github.com/FlashML-org/FreeToken/pull/669"},"FlashML-org/FreeToken#670":{"body_hash":"3b96fdcfbb98a3a4","comments":3,"kind":"pr","number":670,"repo":"FlashML-org/FreeToken","state":"closed","title":"refactor(cpu-moe): split the executor into weight formats and gpu sync backends","updated_at":"2026-10-14T03:12:59Z","url":"https://github.com/FlashML-org/FreeToken/pull/670"},"FlashML-org/FreeToken#671":{"body_hash":"6e6cbc5d8f8fb6c4","comments":0,"kind":"issue","number":671,"repo":"FlashML-org/FreeToken","state":"open","title":"MoE cache and CPU expert refactor","updated_at":"2026-10-14T03:30:45Z","url":"https://github.com/FlashML-org/FreeToken/issues/671"},"FlashML-org/FreeToken#672":{"body_hash":"3912e81c4a31ea89","comments":0,"kind":"pr","number":672,"repo":"FlashML-org/FreeToken","state":"open","title":"feat(kvcache): --kv-placement host keeps the QSA K/V in pinned host memory","updated_at":"2026-10-13T14:06:49Z","url":"https://github.com/FlashML-org/FreeToken/pull/672"},"FlashML-org/FreeToken#673":{"body_hash":"81366ae194333062","comments":0,"kind":"issue","number":673,"repo":"FlashML-org/FreeToken","state":"open","title":"DeepSeek-V4.1's fp4_e4m3_b16 pack rounds a tied scale up on pre-sm_89 GPUs","updated_at":"2026-10-13T14:07:07Z","url":"https://github.com/FlashML-org/FreeToken/issues/673"},"FlashML-org/FreeToken#675":{"body_hash":"5fbe84d3180ec76e","comments":0,"kind":"issue","number":675,"repo":"FlashML-org/FreeToken","state":"open","title":"NVFP4 QSA sparse attention (FlashML-org#408) spills under Triton 3.8: ~10x slower prefill and ~550 MiB local memory","updated_at":"2026-10-13T16:54:30Z","url":"https://github.com/FlashML-org/FreeToken/issues/675"},"FlashML-org/FreeToken#676":{"body_hash":"6ed7131082d34504","comments":2,"kind":"pr","number":676,"repo":"FlashML-org/FreeToken","state":"open","title":"feat(moe): load a short prefill chunk's experts on demand through the slot cache","updated_at":"2026-10-13T19:46:01Z","url":"https://github.com/FlashML-org/FreeToken/pull/676"},"FlashML-org/FreeToken#677":{"body_hash":"5fa2424cec41ab37","comments":0,"kind":"issue","number":677,"repo":"FlashML-org/FreeToken","state":"open","title":"perf(moe): prefetch the next layer's experts from a router lookahead during offloaded decode","updated_at":"2026-10-13T16:54:15Z","url":"https://github.com/FlashML-org/FreeToken/issues/677"},"FlashML-org/FreeToken#678":{"body_hash":"966245138f27565f","comments":0,"kind":"issue","number":678,"repo":"FlashML-org/FreeToken","state":"open","title":"perf(moe): run the shared expert beside the routed-expert miss copy in offloaded decode","updated_at":"2026-10-13T16:54:17Z","url":"https://github.com/FlashML-org/FreeToken/issues/678"},"FlashML-org/FreeToken#680":{"body_hash":"cbddae76f1b1d83d","comments":0,"kind":"pr","number":680,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(moe): admit a wide offloaded decode step's experts 128 at a time","updated_at":"2026-10-13T19:47:46Z","url":"https://github.com/FlashML-org/FreeToken/pull/680"},"FlashML-org/FreeToken#685":{"body_hash":"ec4a752e8e967181","comments":0,"kind":"pr","number":685,"repo":"FlashML-org/FreeToken","state":"closed","title":"perf(moe): split hybrid decode routes inside the ensure kernel","updated_at":"2026-10-14T03:30:35Z","url":"https://github.com/FlashML-org/FreeToken/pull/685"},"FlashML-org/FreeToken#686":{"body_hash":"31a4aa1236ee724d","comments":0,"kind":"pr","number":686,"repo":"FlashML-org/FreeToken","state":"open","title":"perf(qwen4): add opt-in layer-major long prefill","updated_at":"2026-10-14T03:45:17Z","url":"https://github.com/FlashML-org/FreeToken/pull/686"},"FlashML-org/FreeToken#687":{"body_hash":"069318d4717dc42f","comments":0,"kind":"pr","number":687,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(rocm): route moe_align around Triton 3.8 miscompile","updated_at":"2026-10-14T07:13:00Z","url":"https://github.com/FlashML-org/FreeToken/pull/687"},"FlashML-org/FreeToken#688":{"body_hash":"ac26240e1563ff89","comments":0,"kind":"pr","number":688,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(server): opt-in backend recovery after runtime CUDA OOM","updated_at":"2026-10-14T07:40:34Z","url":"https://github.com/FlashML-org/FreeToken/pull/688"}},"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 7009:9A557:6DE528:16A6D2C:6ACF320C and timestamp 2026-10-14 07:41:00 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":"2026-10-14T07:40:32+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":"ROCm Core SDK 10.1.0 Release","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":"62ccb2438bdcb42c","comments":17,"kind":"issue","number":79,"repo":"FlashML-org/FreeToken","state":"open","title":"FreeToken Roadmap (2026)","updated_at":"2026-10-14T03:33:49Z","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":"377a3bc4284f032e9d30777ce099828a158e7e4c","body_hash":"31c9b77d36e57cab","checks":{"failed":[],"pending":[],"state":"pending","total":1},"comments":3,"commits":3,"draft":false,"head_sha":"5309ca655c2c7b21e09bbb91a88b94d51b941b3d","merged":false,"number":133,"review_comments":0,"state":"closed","title":"fix(rocm): make TVM-FFI index and store JIT kernels portable to HIP","updated_at":"2026-10-08T20:32:48Z","url":"https://github.com/FlashML-org/FreeToken/pull/133"},"134":{"base_sha":"377a3bc4284f032e9d30777ce099828a158e7e4c","body_hash":"cf8d7ae82d59f62c","checks":{"failed":[],"pending":[],"state":"pending","total":1},"comments":1,"commits":2,"draft":false,"head_sha":"decc3d4e1337891fc31a45aa8a02fa383f24ae3f","merged":true,"number":134,"review_comments":0,"state":"closed","title":"fix(rocm): gate CUDA-only optional backends on ROCm","updated_at":"2026-10-09T02:55:57Z","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: 2026-10-14T07:40:32+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 (2026-10-06T06:24:26+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.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-07T17:26:12+00:00).

• Новый релевантный pr: rocm: fix native extensions, kernel JIT builds, and Triton PTX fallbacks for gfx1150.
• ⚠️ 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 4803:B9AD:14808F:42752C:6AC680CF and timestamp 2026-10-07 17:26:39 UTC. Fo…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-07T23:07:02+00:00).

• Новый релевантный issue: Upgrade to torch 2.14.1, Triton 3.8 and sglang-kernel 0.4.9.
• Новый релевантный pr: build(deps): move to torch 2.14.1, triton 3.8 and sglang-kernel 0.4.9.
• Новый релевантный issue: Pre-sm70: remaining launch-config fixes on the Qwen3.8-Flash-Next path.
• Новый релевантный issue: CPU MoE: AVX2 int8-activation NVFP4 kernel for CPUs without VNNI.
• Новый релевантный issue: Load-time MXFP8 requantization of bf16 dense, hyper-connection and lm_head weights.
• Новый релевантный issue: QSA host KV tier: give each layer its own slot map (decode at 231K 6.8 -> 15.6 tok/s).
• Новый релевантный pr: ci: auto-label amd issues and prs by title.
• perf(moe): spin-wait the cpu moe pool between layers: обновлён (state closed→closed, comments 2→3).
• FlashML-org/FreeToken:main: 127c8566 → 377a3bc4 — ci: auto-label amd issues and prs by title (ci: auto-label amd issues and prs by title FlashML-org/FreeToken#620).
• ⚠️ 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 8003:22D5D9:34133:3A4C8:6AC6D0B5 and timestamp 2026-10-07 23:07:34 UTC. Fo…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-08T06:08:26+00:00).

• fix(rocm): make TVM-FFI index and store JIT kernels portable to HIP: новый commit 9ec8271a → 5309ca65.
• fix(rocm): make TVM-FFI index and store JIT kernels portable to HIP: draft True → False.
• fix(rocm): make TVM-FFI index and store JIT kernels portable to HIP: CI/checks pending → pending.
• fix(rocm): gate CUDA-only optional backends on ROCm: новый commit c4f4a36e → 59598576.
• fix(rocm): gate CUDA-only optional backends on ROCm: draft True → False.
• fix(rocm): gate CUDA-only optional backends on ROCm: CI/checks pending → pending.
• Новый релевантный pr: perf(kernel): keep 16-byte kv loads in dsv4 sparse attention and unroll gated pool.
• Новый релевантный pr: perf(kernel): read dsv41 indexer key rows as int32 words.
• Новый релевантный pr: perf(moe): avoid intra-op fanout while filling NVFP4 host banks.
• Новый релевантный pr: perf(kernel): avoid layout conversions in nvfp4 marlin decode.
• Новый релевантный pr: perf(kernel): skip masks and out-of-window tiles in triton extend attention.
• Upgrade to torch 2.14.1, Triton 3.8 and sglang-kernel 0.4.9: обновлён (state open→open, comments 0→0).
• FlashML-org/FreeToken:main: 377a3bc4 → 1c3aa0aa — perf(kernel): keep 16-byte kv loads in dsv4 sparse attention and unroll gated pool (perf(kernel): keep 16-byte kv loads in dsv4 sparse attention and unroll gated pool FlashML-org/FreeToken#633).
• Upstream main: релевантные изменения 377a3bc4 → 1c3aa0aa — commits: perf(kernel): avoid layout conversions in nvfp4 marlin decode (perf(kernel): avoid layout conversions in nvfp4 marlin decode FlashML-org/FreeToken#631); perf(kernel): read dsv41 indexer key rows as int32 words (perf(kernel): read dsv41 indexer key rows as int32 words FlashML-org/FreeToken#634); perf(kernel): keep 16-byte kv loads in dsv4 sparse attention and unroll gated pool (perf(kernel): keep 16-byte kv loads in dsv4 sparse attention and unroll gated pool FlashML-org/FreeToken#633) | files: python/freetoken/kernel/triton/dsv4/compress.py, python/freetoken/kernel/triton/dsv4/sparse_attn.py, python/freetoken/kernel/triton/dsv41/indexer.py, python/freetoken/kernel/triton/nvfp4_fused_moe.py, python/freetoken/moe/fused_nvfp4.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:3A33A6:599FCB:1262AB1:6AC73381 and timestamp 2026-10-08 06:09:05 UTC.…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-08T13:33:52+00:00).

• fix(rocm): make TVM-FFI index and store JIT kernels portable to HIP: обновлены описание/review/discussion (comments 1→2, review 0→0).
• FreeToken Roadmap (2026): body edited.
• Новый релевантный issue: Measurements: TP=2 over PCIe on consumer GPUs (no NVLink) - the all-reduce costs 26% of SM time in prefill.
• Новый релевантный pr: fix(kernel): correct jit tensor matcher checks and pdl arch guards.
• Новый релевантный pr: feat(quant): opt-in fp8-block QAT for the lm_head.
• Новый релевантный pr: feat(qwen3_5_moe): tensor-parallel weight sharding for multi-GPU serving.
• Upgrade to torch 2.14.1, Triton 3.8 and sglang-kernel 0.4.9: обновлён (state open→open, comments 0→0).
• perf(kernel): skip masks and out-of-window tiles in triton extend attention: обновлён (state open→closed, comments 0→0).
• FlashML-org/FreeToken:main: 1c3aa0aa → 07813248 — perf(kernel): skip masks and out-of-window tiles in triton extend attention (perf(kernel): skip masks and out-of-window tiles in triton extend attention FlashML-org/FreeToken#632).
• Upstream main: релевантные изменения 1c3aa0aa → 07813248 — commits: perf(kernel): skip masks and out-of-window tiles in triton extend attention (perf(kernel): skip masks and out-of-window tiles in triton extend attention FlashML-org/FreeToken#632) | files: python/freetoken/kernel/triton/attention.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 B456:21738E:BA8FAB:2642402:6AC79BDF and timestamp 2026-10-08 13:34:23 UTC.…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-08T19:57:59+00:00).

• Новый релевантный issue: Live cache rebuild grows MoE before shrinking KV/mamba, so a resize that passes the fit check can OOM.
• Новый релевантный pr: fix(engine): apply shrinking cache resizes before growing ones in live rebuild (Closes #643).
• Новый релевантный issue: Qwen3.8-Flash-Next: transformers 5.16 drops \p{M} from the pre-tokenizer, so combining-mark text gets different ids.
• Новый релевантный issue: /v1/chat/completions returns 400 for the OpenAI developer role.
• Новый релевантный issue: --moe-cache-auto asserts "cache budget too small" instead of falling back to prefill overlap off.
• Новый релевантный pr: glm5_next: support GLM-5.3-Flash (hybrid KDA + DSA, mHC, NVFP4 expert offload).
• Новый релевантный issue: [Bug] Desktop installer fails / breaks when changing installation directory to a custom path (e.g., D: drive).
• ⚠️ 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 9C41:2AB5DB:74647:17A665:6AC7F5F0 and timestamp 2026-10-08 19:58:40 UTC. F…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-09T00:24:21+00:00).

• fix(rocm): make TVM-FFI index and store JIT kernels portable to HIP: state open → closed.
• fix(rocm): make TVM-FFI index and store JIT kernels portable to HIP: обновлены описание/review/discussion (comments 2→3, review 0→0).
• FreeToken Roadmap (2026): body edited.
• fix(kernel): correct jit tensor matcher checks and pdl arch guards: обновлён (state open→closed, comments 1→2).
• FlashML-org/FreeToken:main: 07813248 → 23fc922b — feat(deepseek_v4): serve image input on the V4-Flash-Vision release (feat(deepseek_v4): serve image input on the V4-Flash-Vision release FlashML-org/FreeToken#502).
• Upstream main: релевантные изменения 07813248 → 23fc922b — commits: fix(kernel): correct jit tensor matcher checks and pdl arch guards (fix(kernel): correct jit tensor matcher checks and pdl arch guards FlashML-org/FreeToken#638) | files: python/freetoken/kernel/csrc/jit/index.cu, python/freetoken/kernel/csrc/jit/store.cu.
• ⚠️ 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 DC02:36FC9F:1BE99AC:5CE9883:6AC8345E and timestamp 2026-10-09 00:25:02 UTC…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-09T09:08:09+00:00).

• fix(rocm): gate CUDA-only optional backends on ROCm: новый commit 59598576 → decc3d4e.
• fix(rocm): gate CUDA-only optional backends on ROCm: state open → closed.
• fix(rocm): gate CUDA-only optional backends on ROCm: merged False → True.
• Новый релевантный pr: perf(kernel): avoid the activation layout conversion in fp8 moe decode and read mxfp4 gemv weights as column pairs.
• Новый релевантный pr: test(kernel): cover Triton activation and norm fallbacks.
• feat(models): support deepseek v4.1 with native fp8-fp4 kv storage: обновлён (state open→open, comments 2→3).
• Новый релевантный pr: refactor(moe): extract hybrid decode orchestration.
• FlashML-org/FreeToken:main: 23fc922b → 2d8b9cc1 — perf(kernel): avoid the activation layout conversion in fp8 moe decode and read mxfp4 gemv weights as column pairs (perf(kernel): avoid the activation layout conversion in fp8 moe decode and read mxfp4 gemv weights as column pairs FlashML-org/FreeToken#653).
• Upstream main: релевантные изменения 23fc922b → 2d8b9cc1 — commits: fix(rocm): gate CUDA-only optional backends on ROCm (fix(rocm): gate CUDA-only optional backends on ROCm FlashML-org/FreeToken#134); test(kernel): cover Triton activation and norm fallbacks (test(kernel): cover Triton activation and norm fallbacks FlashML-org/FreeToken#652); perf(kernel): avoid the activation layout conversion in fp8 moe decode and read mxfp4 gemv weights as column pairs (perf(kernel): avoid the activation layout conversion in fp8 moe decode and read mxfp4 gemv weights as column pairs FlashML-org/FreeToken#653) | files: python/freetoken/engine/engine.py, python/freetoken/kernel/backend.py, python/freetoken/kernel/triton/fp8_blockscale_moe.py, python/freetoken/kernel/triton/mxfp4_moe.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 B3C8:34FC5E:1EC68A1:6546BD7:6AC8AF1C and timestamp 2026-10-09 09:08:44 UTC…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-09T17:00:20+00:00).

• Новый релевантный issue: Qwen3.8-27B-NVFP4 fails to load — RuntimeError: CUDA driver error: device not ready.
• ⚠️ 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 3039:1B3078:104B925:353610A:6AC91DD1 and timestamp 2026-10-09 17:01:05 UTC…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-09T22:40:34+00:00).

• Новый релевантный pr: fix(kernel): avoid nan in triton sampling softmax when a chunk starts with -inf.
• ⚠️ 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 900E:3BCE23:331D15:A7878D:6AC96D7C and timestamp 2026-10-09 22:41:00 UTC. …

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-10T05:55:54+00:00).

• Новый релевантный pr: fix(moe): refuse e8m0 scale code 0xFF when MXFP4 expert pieces are read.
• ⚠️ 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 5410:277E2C:A98783:22C6815:6AC9D38A and timestamp 2026-10-10 05:56:26 UTC.…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-10T12:35:26+00:00).

• fix(kernel): avoid nan in triton sampling softmax when a chunk starts with -inf: обновлён (state open→closed, comments 0→0).
• Upgrade to torch 2.14.1, Triton 3.8 and sglang-kernel 0.4.9: обновлён (state open→open, comments 0→0).
• FlashML-org/FreeToken:main: 2d8b9cc1 → ee9789ff — fix(kernel): avoid nan in triton sampling softmax when a chunk starts with -inf (fix(kernel): avoid nan in triton sampling softmax when a chunk starts with -inf FlashML-org/FreeToken#654).
• Upstream main: релевантные изменения 2d8b9cc1 → ee9789ff — commits: fix(kernel): avoid nan in triton sampling softmax when a chunk starts with -inf (fix(kernel): avoid nan in triton sampling softmax when a chunk starts with -inf FlashML-org/FreeToken#654) | files: python/freetoken/kernel/triton/sampling.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 C43B:11A41B:89A68B:1C72338:6ACA312C and timestamp 2026-10-10 12:35:56 UTC.…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-10T18:38:22+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 C416:33B1EA:51E04F:10E957A:6ACA863E and timestamp 2026-10-10 18:38:54 UTC.…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-10T23:34:47+00:00).

• Upgrade to torch 2.14.1, Triton 3.8 and sglang-kernel 0.4.9: обновлён (state open→open, comments 0→0).
• build(deps): move to torch 2.14.1, triton 3.8 and sglang-kernel 0.4.9: обновлён (state open→closed, comments 0→0).
• FlashML-org/FreeToken:main: ee9789ff → 74a38777 — build(deps): move to torch 2.14.1, triton 3.8 and sglang-kernel 0.4.9 (build(deps): move to torch 2.14.1, triton 3.8 and sglang-kernel 0.4.9 FlashML-org/FreeToken#630).
• Upstream main: релевантные изменения ee9789ff → 74a38777 — commits: build(deps): move to torch 2.14.1, triton 3.8 and sglang-kernel 0.4.9 (build(deps): move to torch 2.14.1, triton 3.8 and sglang-kernel 0.4.9 FlashML-org/FreeToken#630) | files: freetoken-kernel-cache/pyproject.toml, pyproject.toml, setup.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 8C03:24529C:85DE889:1B90D520:6ACACBB0 and timestamp 2026-10-10 23:35:12 UT…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-11T05:50:28+00:00).

• Новый релевантный issue: [Bug] Corrupted Triton JIT cache crash-loops the engine: worker dies on kernel load, auto-restart can never heal disk poison.
• ⚠️ 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 D412:3E5351:18C31AC:1D87081:6ACB23C8 and timestamp 2026-10-11 05:51:04 UTC…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-11T12:53:58+00:00).

• Новый релевантный issue: Requests stall and time out above --max-running-requests 8; GPU stays at ~28% of TDP and only one CPU thread is busy.
• Новый релевантный pr: fix(cpu-moe): honor padded fp8 scale strides, add avx512f tier and float64 parity (builds on #36).
• Новый релевантный pr: feat(moe): support gsq q2_0 routed experts.
• Новый релевантный pr: fix(kernel): enable disk row stores on windows.
• ⚠️ 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 EC31:8CE46:32F7409:A832107:6ACB86FE and timestamp 2026-10-11 12:54:22 UTC.…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-11T18:56:45+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 4C3C:3AC91B:7CE7BF:19C57AE:6ACBDC06 and timestamp 2026-10-11 18:57:10 UTC.…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-11T23:41:12+00:00).

• FlashML-org/FreeToken:main: 74a38777 → c8186396 — fix(sampling): apply presence and frequency penalties (fix(sampling): apply presence and frequency penalties FlashML-org/FreeToken#527).
• Upstream main: релевантные изменения 74a38777 → c8186396 — files: python/freetoken/engine/sample.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 1B40:369AE0:369685B:B57926A:6ACC1EB4 and timestamp 2026-10-11 23:41:40 UTC…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-12T06:01:50+00:00).

• Новый релевантный issue: FP8 block-quantized MoE models have no usable kernel under moe_strategy=offload (0.1.3).
• fix(cpu-moe): honor padded fp8 scale strides, add avx512f tier and float64 parity (builds on #36): обновлён (state open→closed, comments 3→4).
• Новый релевантный pr: feat(cpu-moe): serve block-fp8 experts on the cpu and hybrid backends.
• FlashML-org/FreeToken:main: c8186396 → 95ac86f2 — feat(cpu-moe): serve block-fp8 experts on the cpu and hybrid backends (feat(cpu-moe): serve block-fp8 experts on the cpu and hybrid backends FlashML-org/FreeToken#662).
• Upstream main: релевантные изменения c8186396 → 95ac86f2 — commits: feat(cpu-moe): serve block-fp8 experts on the cpu and hybrid backends (feat(cpu-moe): serve block-fp8 experts on the cpu and hybrid backends FlashML-org/FreeToken#662) | files: python/freetoken/moe/bench_profile.py, python/freetoken/moe/benchbw.py, python/freetoken/moe/cpu_executor.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 6FC2:DF87C:71B682E:179D5051:6ACC77ED and timestamp 2026-10-12 06:02:21 UTC…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-12T15:30:32+00:00).

• perf(moe): dequant nvfp4 prefill experts by bit placement and load activations contiguously: обновлён (state open→open, comments 0→1).
• Новый релевантный issue: bench bw verdict misses in the other direction: auto stays on offload (ratio ~1.9), while hybrid decodes 21% / 55% faster (Qwen3.8-Flash-Next NVFP4, RTX 4090 + dual Xeon).
• Новый релевантный pr: fix(rocm): make CPU/Hybrid MoE graph replay safe.
• Новый релевантный issue: AMD Roadmap 2026 Q4-2027 Q1.
• Новый релевантный issue: sm75 (Turing) fork: 35B MoE on an 8 GB RTX 2080 — 350/25 t/s, 64K context.
• ⚠️ 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 D00A:3DE026:18B4BCA:5123B30:6ACCFD33 and timestamp 2026-10-12 15:30:59 UTC…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-12T23:00:52+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 EC11:3AF15C:FFD2A:347F83:6ACD66C7 and timestamp 2026-10-12 23:01:27 UTC. F…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-13T06:49:04+00:00).

• Новый релевантный pr: fix(rocm): verify native cpu moe graph handshake on rocm 10.
• fix(rocm): make CPU/Hybrid MoE graph replay safe: обновлён (state open→open, comments 2→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 4430:16C1B4:3BB5C1E:C44AEE7:6ACDD489 and timestamp 2026-10-13 06:49:46 UTC…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-13T16:01:38+00:00).

• Новый релевантный issue: perf(moe): run the shared expert beside the routed-expert miss copy in offloaded decode.
• Новый релевантный issue: perf(moe): prefetch the next layer's experts from a router lookahead during offloaded decode.
• Новый релевантный pr: feat(kvcache): add nvfp4 kv quantization.
• Новый релевантный pr: feat(moe): load a short prefill chunk's experts on demand through the slot cache.
• Новый релевантный issue: NVFP4 QSA sparse attention (#408) spills under Triton 3.8: ~10x slower prefill and ~550 MiB local memory.
• Новый релевантный issue: DeepSeek-V4.1's fp4_e4m3_b16 pack rounds a tied scale up on pre-sm_89 GPUs.
• Новый релевантный pr: feat(kvcache): --kv-placement host keeps the QSA K/V in pinned host memory.
• fix(server)!: answer /health with 503 until the server accepts requests: обновлён (state open→open, comments 0→0).
• perf(linear): run a single-row bf16 projection as one GEMV: обновлён (state open→open, comments 0→0).
• fix(engine): size memory pools against the VRAM other processes leave free on Windows: 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).
• perf(moe): dequant nvfp4 prefill experts by bit placement and load activations contiguously: обновлён (state open→open, comments 1→2).
• 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).
• fix(server): keep a silent stream alive and notice a client that left: обновлён (state open→open, comments 0→0).
• Upgrade to torch 2.14.1, Triton 3.8 and sglang-kernel 0.4.9: обновлён (state open→closed, comments 0→1).
• Новый релевантный issue: MoE cache and CPU expert refactor.
• Новый релевантный pr: refactor(cpu-moe): split the executor into weight formats and gpu sync backends.
• Новый релевантный pr: build(deps): move to flashlib 0.4.0 and drop the numpy<2.5 cap.
• FlashML-org/FreeToken:main: 95ac86f2 → 9c635a15 — build(deps): move to flashlib 0.4.0 and drop the numpy<2.5 cap (build(deps): move to flashlib 0.4.0 and drop the numpy<2.5 cap FlashML-org/FreeToken#669).
• Upstream main: релевантные изменения 95ac86f2 → 9c635a15 — files: pyproject.toml.
• ⚠️ 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 085F:2B34F1:F7B8FB:340EA26:6ACE560F and timestamp 2026-10-13 16:02:23 UTC.…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-13T21:17:08+00:00).

• Новый релевантный pr: fix(moe): admit a wide offloaded decode step's experts 128 at a time.
• feat(moe): load a short prefill chunk's experts on demand through the slot cache: обновлён (state open→open, comments 0→2).
• Qwen3.8-Flash-Next: transformers 5.16 drops \p{M} from the pre-tokenizer, so combining-mark text gets different ids: обновлён (state open→open, comments 0→1).
• /v1/chat/completions returns 400 for the OpenAI developer role: обновлён (state open→open, comments 0→1).
• Live cache rebuild grows MoE before shrinking KV/mamba, so a resize that passes the fit check can OOM: обновлён (state open→open, comments 1→2).
• --moe-cache-auto asserts "cache budget too small" instead of falling back to prefill overlap off: обновлён (state open→open, comments 0→1).
• Pre-sm70: remaining launch-config fixes on the Qwen3.8-Flash-Next path: обновлён (state open→open, comments 0→1).
• CPU MoE: AVX2 int8-activation NVFP4 kernel for CPUs without VNNI: обновлён (state open→open, comments 0→1).
• Новый релевантный issue: Load-time MXFP8 requantization of bf16 dense, hyper-connection and lm_head weights.
• Новый релевантный issue: QSA host KV tier: give each layer its own slot map (decode at 231K 6.8 -> 15.6 tok/s).
• NVFP4 QSA sparse attention (#408) spills under Triton 3.8: ~10x slower prefill and ~550 MiB local memory: обновлён (state open→open, comments 0→0).
• perf(moe): run the shared expert beside the routed-expert miss copy in offloaded decode: обновлён (state open→open, comments 0→0).
• perf(moe): prefetch the next layer's experts from a router lookahead during offloaded decode: 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 FC02:16626D:A34E62:21125C6:6ACE9FFE and timestamp 2026-10-13 21:17:50 UTC.…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-14T01:14:17+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 501F:1DC131:13FEA5E:4186B91:6ACED793 and timestamp 2026-10-14 01:14:59 UTC…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-14T07:40:32+00:00).

• FreeToken Roadmap (2026): body edited.
• Новый релевантный pr: fix(server): opt-in backend recovery after runtime CUDA OOM.
• Новый релевантный pr: fix(rocm): route moe_align around Triton 3.8 miscompile.
• Новый релевантный pr: perf(qwen4): add opt-in layer-major long prefill.
• MoE cache and CPU expert refactor: обновлён (state open→open, comments 0→0).
• Новый релевантный pr: perf(moe): split hybrid decode routes inside the ensure kernel.
• refactor(cpu-moe): split the executor into weight formats and gpu sync backends: обновлён (state open→closed, comments 0→3).
• FlashML-org/FreeToken:main: 9c635a15 → f02fbb16 — perf(moe): split hybrid decode routes inside the ensure kernel (perf(moe): split hybrid decode routes inside the ensure kernel FlashML-org/FreeToken#685).
• Upstream main: релевантные изменения 9c635a15 → f02fbb16 — commits: refactor(cpu-moe): split the executor into weight formats and gpu sync backends (refactor(cpu-moe): split the executor into weight formats and gpu sync backends FlashML-org/FreeToken#670); perf(moe): split hybrid decode routes inside the ensure kernel (perf(moe): split hybrid decode routes inside the ensure kernel FlashML-org/FreeToken#685) | files: python/freetoken/engine/engine.py, python/freetoken/moe/benchbw.py, python/freetoken/moe/cpu_executor.py, python/freetoken/moe/offload_cache.py, python/freetoken/moe/offload_kernels.py, setup.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 7009:9A557:6DE528:16A6D2C:6ACF320C and timestamp 2026-10-14 07:41:00 UTC. …

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