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, Native Windows is classified as pin-uncapped, so #112's per-layer residency never auto-engages FlashML-org/FreeToken#120, 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 FlashML-org/FreeToken#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 feat(rocm): add RDNA3 and RDNA4 runtime foundation FlashML-org/FreeToken#132: open, head 45417560, checks pending — feat(rocm): add RDNA3 and RDNA4 runtime foundation
• PR fix(rocm): make TVM-FFI index and store JIT kernels portable to HIP FlashML-org/FreeToken#133: open / draft, head 89ad7ab0, checks pending — fix(rocm): make TVM-FFI index and store JIT kernels portable to HIP
• PR fix(rocm): gate CUDA-only optional backends on ROCm FlashML-org/FreeToken#134: open / draft, head 9a523dc8, checks pending — fix(rocm): gate CUDA-only optional backends on ROCm
• PR fix(rocm): route tensor parallel communication through RCCL FlashML-org/FreeToken#135: open / draft, head 9d5e3fc8, checks pending — fix(rocm): route tensor parallel communication through RCCL
• PR feat(rocm): enable native GGUF kernels on ROCm (validated on RDNA4) FlashML-org/FreeToken#136: open / draft, head 49591978, checks pending — feat(rocm): enable native GGUF kernels on RDNA4
• PR feat(rocm): serve on AMD GPUs through the HIP toolchain FlashML-org/FreeToken#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 #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":"2026-09-02T02:43:19Z","message":"fix(hf): download the shards the safetensors index names (FlashML-org#336)","repo":"FlashML-org/FreeToken","sha":"a80b4d308a81986fa086ec173d7faa70ba737b2d","url":"FlashML-org@a80b4d308a81986fa086ec173d7faa70ba737b2d"},"Maxritz/FreeToken-rocm-test:main":{"branch":"main","date":"2026-08-26T08:26:56Z","message":"Merge pull request Maxritz#2 from PialGhosh2233/main","repo":"Maxritz/FreeToken-rocm-test","sha":"89f14790de7bbc69dd61a999ee76b86ba8316f26","url":"Maxritz@89f14790de7bbc69dd61a999ee76b86ba8316f26"},"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#103%22:%7B%22body_hash%22:%225cc83924c09bfe42%22,%22comments%22:3,%22kind%22:%22pr%22,%22number%22:103,%22repo%22:%22FlashML-org/FreeToken%22,%22state%22:%22closed%22,%22title%22:%22feat(kvcache): 8-bit KV cache (q8_0 / fp8_e4m3) behind --kv-cache-dtype","updated_at":"2026-08-30T16:04:58Z","url":"https://github.com/FlashML-org/FreeToken/pull/103"},"FlashML-org/FreeToken#104":{"body_hash":"0ba9970c69c92be9","comments":12,"kind":"pr","number":104,"repo":"FlashML-org/FreeToken","state":"open","title":"feat(models): support TP for qwen3_5_moe","updated_at":"2026-09-02T02:29:00Z","url":"https://github.com/FlashML-org/FreeToken/pull/104"},"FlashML-org/FreeToken#164":{"body_hash":"86f6c0cd6dee9851","comments":1,"kind":"issue","number":164,"repo":"FlashML-org/FreeToken","state":"open","title":"Qwen3.5 MoE modelopt_mixed loader assumes shared experts use routed-expert NVFP4 quantization","updated_at":"2026-08-30T18:59:23Z","url":"https://github.com/FlashML-org/FreeToken/issues/164"},"FlashML-org/FreeToken#208":{"body_hash":"86da2a39c01e2f6d","comments":3,"kind":"pr","number":208,"repo":"FlashML-org/FreeToken","state":"open","title":"qwen3_5_moe: support unsloth/Qwen3.8-27B-NVFP4 (mixed-precision NVFP4+FP8)","updated_at":"2026-09-01T14:54:55Z","url":"https://github.com/FlashML-org/FreeToken/pull/208"},"FlashML-org/FreeToken#217":{"body_hash":"0e0dbaaf50222ee4","comments":3,"kind":"pr","number":217,"repo":"FlashML-org/FreeToken","state":"open","title":"feat(rocm): AMD ROCm support (gfx1100) + Qwen3.5-MoE GGUF loader","updated_at":"2026-09-02T04:27:57Z","url":"https://github.com/FlashML-org/FreeToken/pull/217"},"FlashML-org/FreeToken#227":{"body_hash":"ec213ea17037ea8c","comments":12,"kind":"issue","number":227,"repo":"FlashML-org/FreeToken","state":"open","title":"[Bug]: RuntimeError: Rowwise scaling is not currently supported on your device","updated_at":"2026-09-01T23:59:58Z","url":"https://github.com/FlashML-org/FreeToken/issues/227"},"FlashML-org/FreeToken#238":{"body_hash":"44be9d09584ba9be","comments":2,"kind":"issue","number":238,"repo":"FlashML-org/FreeToken","state":"open","title":"[BUG] qwen3_5 dense: mixed-precision NVFP4 crashes in ct_bf16_fuse (Float8 × BFloat16 promotion)","updated_at":"2026-09-01T11:15:04Z","url":"https://github.com/FlashML-org/FreeToken/issues/238"},"FlashML-org/FreeToken#24":{"body_hash":"04716372ce11a2b6","comments":1,"kind":"pr","number":24,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(cuda): support Turing GPUs","updated_at":"2026-08-30T15:46:40Z","url":"https://github.com/FlashML-org/FreeToken/pull/24"},"FlashML-org/FreeToken#248":{"body_hash":"89bb063d23656e84","comments":1,"kind":"issue","number":248,"repo":"FlashML-org/FreeToken","state":"open","title":"Two blocking bugs in Windows Desktop on consumer Blackwell (RTX 50): FP8 row-wise gate crash + Mistral tokenizer empty Chinese output","updated_at":"2026-08-31T02:20:59Z","url":"https://github.com/FlashML-org/FreeToken/issues/248"},"FlashML-org/FreeToken#260":{"body_hash":"cc73f05b77b626de","comments":5,"kind":"pr","number":260,"repo":"FlashML-org/FreeToken","state":"open","title":"ROCm: validate native gfx1151 serving on Strix Halo","updated_at":"2026-09-01T22:59:52Z","url":"https://github.com/FlashML-org/FreeToken/pull/260"},"FlashML-org/FreeToken#263":{"body_hash":"0892462f890d8743","comments":4,"kind":"issue","number":263,"repo":"FlashML-org/FreeToken","state":"open","title":"[Bug] ft checkpoint fails on Qwen3.6-35B-A3B abliterated NVFP4 checkpoints: Missing MoE expert source layers","updated_at":"2026-08-30T18:46:20Z","url":"https://github.com/FlashML-org/FreeToken/issues/263"},"FlashML-org/FreeToken#268":{"body_hash":"90e395f1fec25670","comments":4,"kind":"pr","number":268,"repo":"FlashML-org/FreeToken","state":"open","title":"feat(kvcache): sub-byte Q4_0 and Q6_0 KV cache quantization","updated_at":"2026-09-01T05:39:25Z","url":"https://github.com/FlashML-org/FreeToken/pull/268"},"FlashML-org/FreeToken#272":{"body_hash":"5e9ec8faf54790cc","comments":1,"kind":"issue","number":272,"repo":"FlashML-org/FreeToken","state":"closed","title":"Feature Request: Support Image Input","updated_at":"2026-08-31T06:01:56Z","url":"https://github.com/FlashML-org/FreeToken/issues/272"},"FlashML-org/FreeToken#274":{"body_hash":"1534544dce8143a3","comments":2,"kind":"issue","number":274,"repo":"FlashML-org/FreeToken","state":"open","title":"FrontendAPI] ERROR Backend supervisor: KeyError: 'model.layers.0.mlp.shared_expert.gate_up_proj.weight_scale'","updated_at":"2026-08-30T19:17:08Z","url":"https://github.com/FlashML-org/FreeToken/issues/274"},"FlashML-org/FreeToken#275":{"body_hash":"2325c00fd601603b","comments":2,"kind":"pr","number":275,"repo":"FlashML-org/FreeToken","state":"open","title":"fix: [BUG] qwen3_5 dense: mixed-precision NVFP4 crashes in ct_bf16_fuse (Float8 × BFloat16 promotion)","updated_at":"2026-09-01T11:59:57Z","url":"https://github.com/FlashML-org/FreeToken/pull/275"},"FlashML-org/FreeToken#279":{"body_hash":"56c437aabe3b6485","comments":3,"kind":"pr","number":279,"repo":"FlashML-org/FreeToken","state":"closed","title":"feat(qwen4-exp): mmap PLE tables from disc","updated_at":"2026-09-01T17:19:33Z","url":"https://github.com/FlashML-org/FreeToken/pull/279"},"FlashML-org/FreeToken#281":{"body_hash":"e31beec51d453c7e","comments":7,"kind":"issue","number":281,"repo":"FlashML-org/FreeToken","state":"open","title":"Proposal: reproducible end-to-end benchmark / regression harness","updated_at":"2026-09-02T03:54:28Z","url":"https://github.com/FlashML-org/FreeToken/issues/281"},"FlashML-org/FreeToken#283":{"body_hash":"c18348b1af14fce5","comments":1,"kind":"pr","number":283,"repo":"FlashML-org/FreeToken","state":"closed","title":"fix(models): load NVFP4 expert banks from checkpoints without a safetensors index","updated_at":"2026-09-02T02:46:59Z","url":"https://github.com/FlashML-org/FreeToken/pull/283"},"FlashML-org/FreeToken#288":{"body_hash":"44c341ac12601f0e","comments":2,"kind":"issue","number":288,"repo":"FlashML-org/FreeToken","state":"closed","title":"[Bug] Windows + RTX 50 (sm_120): RuntimeError: Rowwise scaling is not currently supported on your device during CUDA graph capture — dense FP8 W8A8 path","updated_at":"2026-09-02T00:00:42Z","url":"https://github.com/FlashML-org/FreeToken/issues/288"},"FlashML-org/FreeToken#291":{"body_hash":"c9f6119e7760932e","comments":0,"kind":"issue","number":291,"repo":"FlashML-org/FreeToken","state":"open","title":"[WSL2] H2D copies from pinned memory silently no-op after heavy pinned-alloc churn (driver 591.86, CUDA 13.0 runtime)","updated_at":"2026-08-30T12:49:27Z","url":"https://github.com/FlashML-org/FreeToken/issues/291"},"FlashML-org/FreeToken#292":{"body_hash":"a6bd3d06b28d15df","comments":2,"kind":"pr","number":292,"repo":"FlashML-org/FreeToken","state":"open","title":"feat: support GLM-5.3-Flash NVFP4","updated_at":"2026-08-31T06:47:17Z","url":"https://github.com/FlashML-org/FreeToken/pull/292"},"FlashML-org/FreeToken#293":{"body_hash":"05c8c92e5c45231f","comments":1,"kind":"pr","number":293,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(qwen4_exp): fix nvfp4 ftw weight loading, qsa indexer, and ple tables","updated_at":"2026-08-31T02:09:33Z","url":"https://github.com/FlashML-org/FreeToken/pull/293"},"FlashML-org/FreeToken#294":{"body_hash":"3a6dec8b0cbc36de","comments":1,"kind":"pr","number":294,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(qwen3_5_moe): detect shared-expert quant for modelopt mixed checkpoints","updated_at":"2026-09-01T08:32:59Z","url":"https://github.com/FlashML-org/FreeToken/pull/294"},"FlashML-org/FreeToken#296":{"body_hash":"e3a01040807ea2af","comments":1,"kind":"pr","number":296,"repo":"FlashML-org/FreeToken","state":"open","title":"feat(qwen3_5_moe): support compressed-tensors NVFP4 experts with block-fp8 dense side","updated_at":"2026-09-01T08:33:03Z","url":"https://github.com/FlashML-org/FreeToken/pull/296"},"FlashML-org/FreeToken#298":{"body_hash":"82f5c00b90b8d5d3","comments":0,"kind":"pr","number":298,"repo":"FlashML-org/FreeToken","state":"open","title":"Support poolside/Laguna-S-2.1-NVFP4, and fix serving from source on Windows","updated_at":"2026-08-30T20:25:02Z","url":"https://github.com/FlashML-org/FreeToken/pull/298"},"FlashML-org/FreeToken#299":{"body_hash":"59cf775ec5fee7c1","comments":0,"kind":"issue","number":299,"repo":"FlashML-org/FreeToken","state":"open","title":"FreeToken-Intel: 35B MoE, live tokens off an Arc Pro B70 (SYCL / XPU port)","updated_at":"2026-09-01T23:55:22Z","url":"https://github.com/FlashML-org/FreeToken/issues/299"},"FlashML-org/FreeToken#300":{"body_hash":"b1273dfe00a571e9","comments":0,"kind":"pr","number":300,"repo":"FlashML-org/FreeToken","state":"open","title":"feat: Automatic KV/MoE Laddering for decode speed vs context-length trade off ( upto 33% faster decode )","updated_at":"2026-08-30T22:35:41Z","url":"https://github.com/FlashML-org/FreeToken/pull/300"},"FlashML-org/FreeToken#301":{"body_hash":"40475c88b787fc6e","comments":1,"kind":"issue","number":301,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(server): allow overriding the distributed port","updated_at":"2026-08-31T02:09:36Z","url":"https://github.com/FlashML-org/FreeToken/issues/301"},"FlashML-org/FreeToken#302":{"body_hash":"53ce92fba4ae82bc","comments":1,"kind":"issue","number":302,"repo":"FlashML-org/FreeToken","state":"open","title":"feat(server): add a server-side thinking override","updated_at":"2026-08-31T02:09:37Z","url":"https://github.com/FlashML-org/FreeToken/issues/302"},"FlashML-org/FreeToken#303":{"body_hash":"987029c9a94f59f1","comments":0,"kind":"issue","number":303,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(engine): account for attention workspace in cache-pool sizing","updated_at":"2026-08-30T23:02:07Z","url":"https://github.com/FlashML-org/FreeToken/issues/303"},"FlashML-org/FreeToken#304":{"body_hash":"6947af227ec8c861","comments":0,"kind":"issue","number":304,"repo":"FlashML-org/FreeToken","state":"open","title":"[Windows] Qwen3.6-35B-A3B-NVFP4 expert loading crashes — torch_cpu.dll access violation (0xc0000005)","updated_at":"2026-08-31T00:47:27Z","url":"https://github.com/FlashML-org/FreeToken/issues/304"},"FlashML-org/FreeToken#306":{"body_hash":"5f4417a782901609","comments":0,"kind":"issue","number":306,"repo":"FlashML-org/FreeToken","state":"open","title":"Long-context decode: client wall-clock is ~half the scheduler gen throughput","updated_at":"2026-08-31T03:09:53Z","url":"https://github.com/FlashML-org/FreeToken/issues/306"},"FlashML-org/FreeToken#307":{"body_hash":"0c7eb3672129dec5","comments":0,"kind":"issue","number":307,"repo":"FlashML-org/FreeToken","state":"open","title":"temperature 0 is not reproducible with thinking on (identical requests diverge on near-ties)","updated_at":"2026-08-31T03:09:53Z","url":"https://github.com/FlashML-org/FreeToken/issues/307"},"FlashML-org/FreeToken#308":{"body_hash":"257cd42bf835ad1d","comments":0,"kind":"issue","number":308,"repo":"FlashML-org/FreeToken","state":"open","title":"ft bench bw -o writes a profile that hybrid never loads (silent fetch cap of 1)","updated_at":"2026-08-31T03:09:54Z","url":"https://github.com/FlashML-org/FreeToken/issues/308"},"FlashML-org/FreeToken#311":{"body_hash":"05fb3389b262f3c4","comments":2,"kind":"pr","number":311,"repo":"FlashML-org/FreeToken","state":"closed","title":"feat(qwen4_exp): stream the PLE n-gram table from disk","updated_at":"2026-09-01T20:35:14Z","url":"https://github.com/FlashML-org/FreeToken/pull/311"},"FlashML-org/FreeToken#315":{"body_hash":"4bb9eb5f4d845e71","comments":0,"kind":"pr","number":315,"repo":"FlashML-org/FreeToken","state":"open","title":"Fix GGUF user-defined token atomicity","updated_at":"2026-08-31T19:08:05Z","url":"https://github.com/FlashML-org/FreeToken/pull/315"},"FlashML-org/FreeToken#316":{"body_hash":"97a5efc93019df65","comments":0,"kind":"pr","number":316,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(rocm): avoid indirect host pointers during HIP graph capture","updated_at":"2026-09-01T15:05:05Z","url":"https://github.com/FlashML-org/FreeToken/pull/316"},"FlashML-org/FreeToken#318":{"body_hash":"b288fbd200e414b8","comments":1,"kind":"issue","number":318,"repo":"FlashML-org/FreeToken","state":"open","title":"Design discussion: first-class out-of-tree model providers","updated_at":"2026-09-01T15:05:50Z","url":"https://github.com/FlashML-org/FreeToken/issues/318"},"FlashML-org/FreeToken#320":{"body_hash":"dc06711149928e6d","comments":0,"kind":"pr","number":320,"repo":"FlashML-org/FreeToken","state":"open","title":"qwen4_exp: load modelopt MIXED_PRECISION (NVFP4 experts + block-FP8 dense) checkpoints","updated_at":"2026-09-01T03:34:03Z","url":"https://github.com/FlashML-org/FreeToken/pull/320"},"FlashML-org/FreeToken#324":{"body_hash":"644fad6946baa855","comments":2,"kind":"issue","number":324,"repo":"FlashML-org/FreeToken","state":"open","title":"[Showcase & Fix] Qwen3.8-Flash-Next on Dual RTX 5090 Blackwell + Tool Call Recovery Patch","updated_at":"2026-09-01T19:44:50Z","url":"https://github.com/FlashML-org/FreeToken/issues/324"},"FlashML-org/FreeToken#328":{"body_hash":"d74f7aedfdc8a045","comments":0,"kind":"issue","number":328,"repo":"FlashML-org/FreeToken","state":"open","title":"Can't load Qwen3.8-Flash-Next","updated_at":"2026-09-01T08:57:35Z","url":"https://github.com/FlashML-org/FreeToken/issues/328"},"FlashML-org/FreeToken#332":{"body_hash":"820355a82f7eb2c1","comments":0,"kind":"pr","number":332,"repo":"FlashML-org/FreeToken","state":"closed","title":"feat(glm5_next): support GLM-5.3-Flash","updated_at":"2026-09-01T22:18:16Z","url":"https://github.com/FlashML-org/FreeToken/pull/332"},"FlashML-org/FreeToken#335":{"body_hash":"42a24e63af908edf","comments":0,"kind":"issue","number":335,"repo":"FlashML-org/FreeToken","state":"open","title":"[Bug] b12x NVFP4 backend on sm_120 (RTX 5090) serves batch 1 only: not CUDA-graph capturable, and the worker dies with a tile-config ValueError at batch >= 2 (Qwen3.8-Flash-Next-NVFP4)","updated_at":"2026-09-01T23:08:23Z","url":"https://github.com/FlashML-org/FreeToken/issues/335"},"FlashML-org/FreeToken#337":{"body_hash":"1b7d3c219882248c","comments":0,"kind":"pr","number":337,"repo":"FlashML-org/FreeToken","state":"open","title":"feat(moe): NVMe disk tier for MoE expert banks","updated_at":"2026-09-02T03:52:50Z","url":"https://github.com/FlashML-org/FreeToken/pull/337"},"FlashML-org/FreeToken#338":{"body_hash":"2ea2d570a433b53e","comments":0,"kind":"pr","number":338,"repo":"FlashML-org/FreeToken","state":"open","title":"perf(ple): fuse the n-gram row-id hash into one Triton kernel","updated_at":"2026-09-02T04:12:47Z","url":"https://github.com/FlashML-org/FreeToken/pull/338"},"FlashML-org/FreeToken#339":{"body_hash":"7d0f17426b165cb7","comments":0,"kind":"pr","number":339,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(kernels): drop the D2H sync from the varlen GDN/KDA prefill conv","updated_at":"2026-09-02T04:12:49Z","url":"https://github.com/FlashML-org/FreeToken/pull/339"},"FlashML-org/FreeToken#340":{"body_hash":"4c2febbad30c626f","comments":0,"kind":"pr","number":340,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(engine): reserve the pool's dummy page in the --moe-cache-auto KV floor","updated_at":"2026-09-02T04:12:51Z","url":"https://github.com/FlashML-org/FreeToken/pull/340"},"FlashML-org/FreeToken#55":{"body_hash":"049b3cf948a4f89e","comments":0,"kind":"issue","number":55,"repo":"FlashML-org/FreeToken","state":"closed","title":"Windows/WDDM: offload serving fails when bank sources exceed the ~50%-RAM pageable-locking pool (repro, measurements, proposed path)","updated_at":"2026-08-31T06:50:04Z","url":"https://github.com/FlashML-org/FreeToken/issues/55"},"FlashML-org/FreeToken#60":{"body_hash":"9ff8ed296bc092d9","comments":17,"kind":"issue","number":60,"repo":"FlashML-org/FreeToken","state":"open","title":"Feature request: AMD GPU support","updated_at":"2026-09-01T23:13:18Z","url":"https://github.com/FlashML-org/FreeToken/issues/60"}},"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 5400:168969:2614347:276478E:6A97A9F1 and timestamp 2026-09-02 04:45:37 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":4,"kind":"issue","number":6461,"relevance":"high","relevance_reason":"Windows + gfx1201 + hipBLASLt/rocBLAS sequence/state-dependent GEMM failure","repo":"ROCm/legacy-rocm-build","state":"open","title":"[Issue]: rocBLAS bf16/f16 GEMM fails with hipErrorInvalidValue on 2nd call, gfx1201 / Windows / ROCm 7.14","updated_at":"2026-08-25T10:25:55Z","url":"https://github.com/ROCm/legacy-rocm-build/issues/6461"}},"last_checked":"2026-09-02T04:45:03+00:00","releases":{"FlashML-org/FreeToken":{"draft":false,"name":"v0.1.2","prerelease":false,"published_at":"2026-08-19T06:20:18Z","tag":"v0.1.2","url":"https://github.com/FlashML-org/FreeToken/releases/tag/v0.1.2"},"ROCm/ROCm":{"draft":false,"name":"ROCm 7.14.0 Release","prerelease":false,"published_at":"2026-07-16T03:28:58Z","tag":"rocm-7.14.0","url":"https://github.com/ROCm/legacy-rocm-build/releases/tag/rocm-7.14.0"},"ROCm/TheRock":{"draft":false,"name":"","prerelease":false,"published_at":"2026-08-26T10:03:34Z","tag":"therock-10.0","url":"https://github.com/ROCm/TheRock/releases/tag/therock-10.0"},"apache/tvm-ffi":{"draft":false,"name":"v0.1.13-post3","prerelease":false,"published_at":"2026-08-10T13:31:09Z","tag":"v0.1.13-post3","url":"https://github.com/apache/tvm-ffi/releases/tag/v0.1.13-post3"},"ggml-org/llama.cpp":{"draft":false,"name":"v0.3.0","prerelease":false,"published_at":"2026-08-25T10:22:58Z","tag":"v0.3.0","url":"https://github.com/ggml-org/llama.cpp/releases/tag/v0.3.0"},"triton-lang/triton-windows":{"draft":false,"name":"v3.8.0-windows.post28","prerelease":false,"published_at":"2026-08-29T12:24:14Z","tag":"v3.8.0-windows.post28","url":"https://github.com/triton-lang/triton-windows/releases/tag/v3.8.0-windows.post28"}},"tracked_issues":{"110":{"body_hash":"d6672f16a90f2854","comments":1,"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":"2026-08-23T20:43:34Z","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":"2fe32010327776c6","comments":15,"kind":"issue","number":79,"repo":"FlashML-org/FreeToken","state":"open","title":"FreeToken Roadmap (2026)","updated_at":"2026-09-01T23:28:49Z","url":"https://github.com/FlashML-org/FreeToken/issues/79"},"82":{"body_hash":"ff59a7ff00722b01","comments":16,"kind":"issue","number":82,"repo":"FlashML-org/FreeToken","state":"open","title":"ROCm RDNA4 gfx1201 Ported. Windows only for now. Test please","updated_at":"2026-08-28T14:46:41Z","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":"bd372b630a028e3faa51f4ab0ef6a98c2f2de501","body_hash":"f5b92547538a8571","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":4,"commits":9,"draft":false,"head_sha":"c0713e44eace3cacd02f27f13d0f6032f4bc86ae","merged":false,"number":132,"review_comments":0,"state":"open","title":"feat(rocm): add RDNA3 and RDNA4 runtime foundation","updated_at":"2026-09-01T05:20:37Z","url":"https://github.com/FlashML-org/FreeToken/pull/132"},"133":{"base_sha":"3a20a79038338c33bd051c52152e6d1faa4d9791","body_hash":"eb17018abe601f2c","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":0,"commits":11,"draft":true,"head_sha":"07a352a0ef5c37398f13ef893648dbbfd126f601","merged":false,"number":133,"review_comments":0,"state":"open","title":"fix(rocm): make TVM-FFI index and store JIT kernels portable to HIP","updated_at":"2026-08-31T08:13:20Z","url":"https://github.com/FlashML-org/FreeToken/pull/133"},"134":{"base_sha":"3a20a79038338c33bd051c52152e6d1faa4d9791","body_hash":"467c9a375cece285","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":0,"commits":11,"draft":true,"head_sha":"32dfb7bd123d935e256508fbc8c76bbaa9012226","merged":false,"number":134,"review_comments":0,"state":"open","title":"fix(rocm): gate CUDA-only optional backends on ROCm","updated_at":"2026-08-31T08:13:20Z","url":"https://github.com/FlashML-org/FreeToken/pull/134"},"135":{"base_sha":"3a20a79038338c33bd051c52152e6d1faa4d9791","body_hash":"3a0be6c359e61828","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":0,"commits":11,"draft":true,"head_sha":"b794aa50056ef3c7a567e02f4787cab3d5b750a9","merged":false,"number":135,"review_comments":0,"state":"open","title":"fix(rocm): route tensor parallel communication through RCCL","updated_at":"2026-08-31T08:13:20Z","url":"https://github.com/FlashML-org/FreeToken/pull/135"},"136":{"base_sha":"3a20a79038338c33bd051c52152e6d1faa4d9791","body_hash":"c875bbbd1cb60d21","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":1,"commits":13,"draft":true,"head_sha":"7b3acf2f3d3d321ba7127d65168f9882d2a5d76f","merged":false,"number":136,"review_comments":0,"state":"open","title":"feat(rocm): enable native GGUF kernels on ROCm (validated on RDNA4)","updated_at":"2026-08-31T08:13:20Z","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: 2026-09-02T04:45:03+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 FreeToken ROCm Watch — notification channel #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 Adding Developer Contributions from Pial Ghosh Maxritz/FreeToken-ROCm#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 (2026-08-26T17:33:54+00:00).

• FreeToken Roadmap (2026): comments 7→8.
• Новый релевантный pr: feat(rocm): AMD ROCm support (gfx1100) + Qwen3.5-MoE GGUF loader.
• ⚠️ 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 3C34:1A0AD0:3D5C1B3:CC983FC:6A8F239F and timestamp 2026-08-26 17:34:23 UTC…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-26T20:02:52+00:00).

• FreeToken Roadmap (2026): comments 8→10.
• Новый релевантный issue: RTX 4070 Ti Super (SM89): every request hangs ~123s then worker dies (empty response) — Qwen3.6-35B-A3B-NVFP4 offload.
• Новый релевантный issue: [Feature Request] Model architecture Qwen3VLForConditionalGeneration not supported.
• ⚠️ 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 0819:391584:A5DD1:C8AB7:6A8F4690 and timestamp 2026-08-26 20:03:28 UTC. Fo…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-26T23:36:03+00:00).

• FreeToken Roadmap (2026): body edited.
• RTX 4070 Ti Super (SM89): every request hangs ~123s then worker dies (empty response) — Qwen3.6-35B-A3B-NVFP4 offload: обновлён (state open→open, comments 0→0).
• Новый релевантный issue: [Bug]: RuntimeError: Rowwise scaling is not currently supported on your device.
• feat(rocm): AMD ROCm support (gfx1100) + Qwen3.5-MoE GGUF loader: обновлён (state open→open, comments 0→1).
• Feature request: AMD GPU support: обновлён (state open→open, comments 12→13).
• ⚠️ 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 D019:3AD3E4:190576:1E7853:6A8F7883 and timestamp 2026-08-26 23:36:35 UTC. …

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-27T09:52:05+00:00).

• FreeToken Roadmap (2026): comments 10→11.
• ROCm RDNA4 gfx1201 Ported. Windows only for now. Test please: comments 9→12.
• 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 1→6, body edited.
• Новый релевантный issue: [Ecosystem Integration] Proposal: Connecting WMP-compliant OS Maintenance Agents via FreeToken Local MoE Streaming.
• Новый релевантный pr: feat(qwen4): support Qwen3.8 Flash Next NVFP4, vision, and Windows serving.
• ⚠️ 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 C843:1EBA7D:11DBAD4:3C714D1:6A9008E8 and timestamp 2026-08-27 09:52:40 UTC…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-27T20:32:35+00:00).

• FreeToken Roadmap (2026): comments 11→14.
• Новый релевантный pr: rocm: fix native extensions, kernel JIT builds, and Triton PTX fallbacks for gfx1150.
• [Bug]: RuntimeError: Rowwise scaling is not currently supported on your device: обновлён (state open→open, comments 1→2).
• Новый релевантный issue: RTX 2050 4GB: Qwen3.6-35B-A3B-NVFP4 runs at ~21 tok/s with single-layer CPU MoE prefill buffer.
• Новый релевантный issue: Qwen3.8-27B-FP8 fails during fp8 weight load with 'CUDA driver error: device not ready' (Qwen3.6-35B-A3B-FP8 and gpt-oss-120b load fine in the same env).
• feat(models): support TP for qwen3_5_moe: обновлён (state open→open, comments 8→9).
• Новый релевантный issue: Prefill hangs on RTX 4060 Ti (SM89) for prompts beyond warmup length.
• Feature: Predictive Expert Offloading: обновлён (state open→open, comments 0→1).
• Новый релевантный issue: [BUG] qwen3_5 dense: mixed-precision NVFP4 crashes in ct_bf16_fuse (Float8 × BFloat16 promotion).
• Новый релевантный issue: Qwen3.6-35B-A3B-Uncensored-NVFP4.
• ⚠️ 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 CBC6:33907B:5307380:11794EC4:6A909F0C and timestamp 2026-08-27 20:33:16 UT…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-28T06:08:28+00:00).

• FreeToken Roadmap (2026): comments 14→15.
• ROCm RDNA4 gfx1201 Ported. Windows only for now. Test please: comments 12→14.
• feat(qwen4): support Qwen3.8 Flash Next NVFP4, vision, and Windows serving: обновлён (state open→open, comments 0→4).
• Новый релевантный pr: fix(kernel): avoid row-wise _scaled_mm stall on sm_89 with torch<2.12.
• Bug in PyTorch 2.11 causes SM89 hangs: обновлён (state open→open, comments 0→1).
• [Bug]: RuntimeError: Rowwise scaling is not currently supported on your device: обновлён (state open→open, comments 2→5).
• Новый релевантный issue: NotImplementedError: GPUs on both sides of the sm_89 fp8 boundary in one process: the host-side e4m3 convention is process-global.
• Новый релевантный issue: [BUG] 0.2.0-beta.15 - Linux, KDE Plasma: Entire GUI "Locks Up" / Freezes when clicking on "Chat".
• [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 1→2).
• rocm: fix native extensions, kernel JIT builds, and Triton PTX fallbacks for gfx1150: обновлён (state open→open, comments 1→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 5404:2C07F1:9C7D2C:20AC8FB:6A9125F5 and timestamp 2026-08-28 06:08:53 UTC.…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-28T18:44:42+00:00).

• FreeToken Roadmap (2026): body edited.
• ROCm RDNA4 gfx1201 Ported. Windows only for now. Test please: comments 14→16.
• Новый релевантный issue: --moe-backend offload hangs indefinitely on first inference request (zero bytes returned); default hybrid backend works fine.
• Новый релевантный issue: Fails to load qwen3.6 or recognize dual GPUs.
• Новый релевантный pr: feat(qwen4_exp): support Qwen3.8-Flash-Next.
• Новый релевантный pr: feat(bench): pass ft serve options through '--' in bench_decode_moe.
• Новый релевантный pr: feat(sm70): support Volta / Tesla V100 via torch 2.10 cu128 downgrade.
• Новый релевантный issue: [Feature Request] Support compressed-tensors per-channel FP8 checkpoints for Qwen3.5 MoE.
• RTX 2050 4GB: Qwen3.6-35B-A3B-NVFP4 runs at ~21 tok/s with single-layer CPU MoE prefill buffer: обновлён (state open→open, comments 1→2).
• Новый релевантный issue: Convert and destroy...
• Новый релевантный issue: Two blocking bugs in Windows Desktop on consumer Blackwell (RTX 50): FP8 row-wise gate crash + Mistral tokenizer empty Chinese output.
• Qwen3.8-27B-FP8 fails during fp8 weight load with 'CUDA driver error: device not ready' (Qwen3.6-35B-A3B-FP8 and gpt-oss-120b load fine in the same env): обновлён (state open→open, comments 0→1).
• feat(qwen4): support Qwen3.8 Flash Next NVFP4, vision, and Windows serving: обновлён (state open→open, comments 4→5).
• ⚠️ 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 E08D:198ECE:2D16A0:97572B:6A91D73A and timestamp 2026-08-28 18:45:14 UTC. …

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-29T02:49:54+00:00).

• FreeToken Roadmap (2026): body edited.
• Новый релевантный pr: ROCm: validate native gfx1151 serving on Strix Halo.
• --moe-backend offload hangs indefinitely on first inference request (zero bytes returned); default hybrid backend works fine: обновлён (state open→closed, comments 0→1).
• fix(kernel): avoid row-wise _scaled_mm stall on sm_89 with torch<2.12: обновлён (state open→open, comments 0→1).
• qwen3_5_moe: support unsloth/Qwen3.8-27B-NVFP4 (mixed-precision NVFP4+FP8): обновлён (state open→open, comments 0→0).
• Новый релевантный issue: [Bug] ft checkpoint fails on Qwen3.6-35B-A3B abliterated NVFP4 checkpoints: Missing MoE expert source layers.
• [Bug]: RuntimeError: Rowwise scaling is not currently supported on your device: обновлён (state open→open, comments 5→6).
• feat(qwen4_exp): support Qwen3.8-Flash-Next: обновлён (state open→closed, comments 0→3).
• Новый релевантный issue: ft serve hangs at 96% weight load on Threadripper 1950X + RTX 5060 Ti.
• FlashML-org/FreeToken:main: 9ef36513 → a05c2654 — docs(models): add Qwen3.8-Flash-Next & Qwen3.8 27b.
• Upstream main: релевантные изменения 9ef36513 → a05c2654 — files: python/freetoken/engine/engine.py, python/freetoken/kernel/triton/hc.py, python/freetoken/kernel/triton/moe_router.py, python/freetoken/kernel/triton/moe_shared_gate.py, python/freetoken/kernel/triton/ple.py, python/freetoken/kernel/triton/qsa/init.py, python/freetoken/kernel/triton/qsa/attend.py, python/freetoken/kernel/triton/qsa/compress.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 AC25:1B56D3:1E6B585:64B0705:6A9248F0 and timestamp 2026-08-29 02:50:24 UTC…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-29T11:07:25+00:00).

• Request hang (silent, zero-log) with hybrid_radix cache + moe-backend offload + nvfp4 triton backend — reproduced on 2 independent models, not explained by #110: state open→closed, comments 0→1.
• Новый релевантный pr: glm5_next: support GLM-5.3-Flash (hybrid KDA + DSA, mHC, NVFP4 expert offload).
• Новый релевантный pr: feat(kvcache): sub-byte Q4_0 and Q6_0 KV cache quantization.
• feat(rocm): AMD ROCm support (gfx1100) + Qwen3.5-MoE GGUF loader: обновлён (state open→open, comments 1→2).
• [Bug]: RuntimeError: Rowwise scaling is not currently supported on your device: обновлён (state open→open, comments 6→7).
• feat(qwen4): support Qwen3.8 Flash Next NVFP4, vision, and Windows serving: обновлён (state open→closed, comments 5→6).
• Prefill hangs on RTX 4060 Ti (SM89) for prompts beyond warmup length: обновлён (state open→closed, comments 4→5).
• fix(deps): use PyTorch 2.13 to fix SM89 hangs: обновлён (state open→closed, comments 0→1).
• Bug in PyTorch 2.11 causes SM89 hangs: обновлён (state open→closed, comments 1→1).
• fix(kernel): avoid row-wise _scaled_mm stall on sm_89 with torch<2.12: обновлён (state open→closed, comments 1→3).
• FlashML-org/FreeToken:main: a05c2654 → 58f4b9ec — fix(kernel): avoid row-wise _scaled_mm stall on sm_89 with torch<2.12 (fix(kernel): avoid row-wise _scaled_mm stall on sm_89 with torch<2.12 FlashML-org/FreeToken#243).
• Upstream main: релевантные изменения a05c2654 → 58f4b9ec — commits: fix(kernel): avoid row-wise _scaled_mm stall on sm_89 with torch<2.12 (fix(kernel): avoid row-wise _scaled_mm stall on sm_89 with torch<2.12 FlashML-org/FreeToken#243) | files: python/freetoken/kernel/triton/fp8_pertensor_linear.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 33C4:1FED24:313D4D2:A51B4C9:6A92BD88 and timestamp 2026-08-29 11:07:52 UTC…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-29T16:44:19+00:00).

• Новый релевантный pr: feat(qwen4-exp): mmap PLE tables from disc.
• Новый релевантный pr: fix(moe): match the benchbw verdict to the served model's expert size.
• Новый релевантный issue: NVFP4 unusable on sm_120 workstation Blackwell (RTX PRO 4000 SFF): "Rowwise scaling is not currently supported on your device".
• [Bug]: RuntimeError: Rowwise scaling is not currently supported on your device: обновлён (state open→open, comments 7→8).
• Новый релевантный issue: Feature Request: Support Image Input.
• [Bug] ft checkpoint fails on Qwen3.6-35B-A3B abliterated NVFP4 checkpoints: Missing MoE expert source layers: обновлён (state open→open, comments 0→2).
• feat(kvcache): sub-byte Q4_0 and Q6_0 KV cache quantization: обновлён (state open→open, comments 0→1).
• Новый релевантный pr: fix: [BUG] qwen3_5 dense: mixed-precision NVFP4 crashes in ct_bf16_fuse (Float8 × BFloat16 promotion).
• Новый релевантный issue: [FrontendAPI] ERROR Backend supervisor: KeyError: 'model.layers.0.mlp.shared_expert.gate_up_proj.weight_scale'](FrontendAPI] ERROR Backend supervisor: KeyError: 'model.layers.0.mlp.shared_expert.gate_up_proj.weight_scale' FlashML-org/FreeToken#274).
• qwen3_5_moe: support unsloth/Qwen3.8-27B-NVFP4 (mixed-precision NVFP4+FP8): обновлён (state open→open, comments 0→1).
• Новый релевантный issue: [Feature Request] Please add SSD offloading for KV cache to save GPU memory (like LM-Cache).
• Новый релиз: triton-lang/triton-windows v3.8.0-windows.post28.
• ⚠️ 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 B822:3C8331:1461DA1:44128E2:6A930C7B and timestamp 2026-08-29 16:44:43 UTC…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-29T20:56:35+00:00).

• Новый релевантный issue: Feature Request: opt-in FP8 KV-cache support for long-context MoE inference.
• ⚠️ 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 A811:3CE456:2509012:7BC17A6:6A9347A3 and timestamp 2026-08-29 20:57:07 UTC…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-30T00:21:59+00:00).

• GGUF: support all quant types, add qwen35moe: обновлены описание/review/discussion (comments 7→8, review 0→0).
• Новый релевантный issue: Model loading fails with OSError: [Errno 22] Invalid argument when safetensors are stored on eCryptFS.
• Новый релевантный issue: Proposal: reproducible end-to-end benchmark / regression harness.
• Новый релевантный pr: fix(cuda): support Turing GPUs.
• ⚠️ 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 F506:280C77:81D704F:1AF2DDD2:6A9377C7 and timestamp 2026-08-30 00:22:31 UT…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-30T08:20:45+00:00).

• ROCm: validate native gfx1151 serving on Strix Halo: обновлён (state open→open, comments 0→1).
• Новый релевантный issue: docs: docs/models.md has no memory column — the expert pool ranges 9.6 GB to 363.8 GB across the listed checkpoints.
• Новый релевантный pr: fix(scheduler): hybrid radix hit corruption — GDN snapshot copy-on-donate + admission barriers.
• feat(kvcache): sub-byte Q4_0 and Q6_0 KV cache quantization: обновлён (state open→open, comments 1→2).
• Новый релевантный issue: [Bug] Windows + RTX 50 (sm_120): RuntimeError: Rowwise scaling is not currently supported on your device during CUDA graph capture — dense FP8 W8A8 path.
• feat(qwen4-exp): mmap PLE tables from disc: обновлён (state open→open, comments 0→1).
• Новый релевантный pr: fix(models): load NVFP4 expert banks from checkpoints without a safetensors index.
• FlashML-org/FreeToken:main: 58f4b9ec → 4b94bdc3 — docs: add SECURITY.md.
• ⚠️ 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 4018:E88B1:4CA8D93:100A4F0B:6A93E7F7 and timestamp 2026-08-30 08:21:11 UTC…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-30T14:53:16+00:00).

• ROCm: validate native gfx1151 serving on Strix Halo: обновлён (state open→open, comments 1→4).
• Новый релевантный issue: [WSL2] H2D copies from pinned memory silently no-op after heavy pinned-alloc churn (driver 591.86, CUDA 13.0 runtime).
• [Bug] ft checkpoint fails on Qwen3.6-35B-A3B abliterated NVFP4 checkpoints: Missing MoE expert source layers: 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 A422:1A3C35:89A40B:D80F6C:6A9443FE and timestamp 2026-08-30 14:53:50 UTC. …

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-30T19:19:43+00:00).

• Новый релевантный pr: feat: support GLM-5.3-Flash NVFP4.
• [FrontendAPI] ERROR Backend supervisor: KeyError: 'model.layers.0.mlp.shared_expert.gate_up_proj.weight_scale'](FrontendAPI] ERROR Backend supervisor: KeyError: 'model.layers.0.mlp.shared_expert.gate_up_proj.weight_scale' FlashML-org/FreeToken#274): обновлён (state open→open, comments 0→2).
• Новый релевантный issue: Qwen3.5 MoE modelopt_mixed loader assumes shared experts use routed-expert NVFP4 quantization.
• Новый релевантный pr: feat(qwen3_5_moe): support compressed-tensors NVFP4 experts with block-fp8 dense side.
• Новый релевантный pr: fix(qwen3_5_moe): detect shared-expert quant for modelopt mixed checkpoints.
• [Bug] Windows + RTX 50 (sm_120): RuntimeError: Rowwise scaling is not currently supported on your device during CUDA graph capture — dense FP8 W8A8 path: обновлён (state open→open, comments 0→1).
• [Bug] ft checkpoint fails on Qwen3.6-35B-A3B abliterated NVFP4 checkpoints: Missing MoE expert source layers: обновлён (state open→open, comments 3→4).
• ROCm: validate native gfx1151 serving on Strix Halo: обновлён (state open→open, comments 4→5).
• Новый релевантный pr: feat(kvcache): 8-bit KV cache (q8_0 / fp8_e4m3) behind --kv-cache-dtype.
• Новый релевантный pr: fix(qwen4_exp): fix nvfp4 ftw weight loading, qsa indexer, and ple tables.
• fix(cuda): support Turing GPUs: обновлён (state open→open, comments 2→1).
• feat(qwen4-exp): mmap PLE tables from disc: 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 9011:1B84E8:175C31B:4F24CB7:6A948279 and timestamp 2026-08-30 19:20:25 UTC…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-30T22:45:54+00:00).

• Новый релевантный pr: feat: Automatic KV/MoE Laddering for decode speed vs context-length trade off ( upto 33% faster decode ).
• Новый релевантный issue: FreeToken-Intel: 35B MoE, live tokens off an Arc Pro B70 (SYCL / XPU port).
• Новый релевантный pr: Support poolside/Laguna-S-2.1-NVFP4, and fix serving from source on Windows.
• feat: support GLM-5.3-Flash NVFP4: 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 4001:17F718:F22A80F:322A5191:6A94B2BA and timestamp 2026-08-30 22:46:18 UT…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-31T05:38:54+00:00).

• 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 2→3.
• feat(kvcache): sub-byte Q4_0 and Q6_0 KV cache quantization: обновлён (state open→open, comments 2→3).
• Новый релевантный issue: ft bench bw -o writes a profile that hybrid never loads (silent fetch cap of 1).
• Новый релевантный issue: temperature 0 is not reproducible with thinking on (identical requests diverge on near-ties).
• Новый релевантный issue: Long-context decode: client wall-clock is ~half the scheduler gen throughput.
• Proposal: reproducible end-to-end benchmark / regression harness: обновлён (state open→open, comments 1→3).
• Two blocking bugs in Windows Desktop on consumer Blackwell (RTX 50): FP8 row-wise gate crash + Mistral tokenizer empty Chinese output: обновлён (state open→open, comments 0→1).
• Новый релевантный issue: feat(server): add a server-side thinking override.
• Новый релевантный issue: fix(server): allow overriding the distributed port.
• fix(qwen4_exp): fix nvfp4 ftw weight loading, qsa indexer, and ple tables: обновлён (state open→open, comments 0→1).
• Новый релевантный issue: [Windows] Qwen3.6-35B-A3B-NVFP4 expert loading crashes — torch_cpu.dll access violation (0xc0000005).
• feat: support GLM-5.3-Flash NVFP4: обновлён (state open→open, comments 1→2).
• Новый релевантный issue: fix(engine): account for attention workspace in cache-pool sizing.
• [Bug]: RuntimeError: Rowwise scaling is not currently supported on your device: обновлён (state open→open, comments 8→9).
• ⚠️ 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 FFC2:3874F9:748001C:7B1798B:6A95138F and timestamp 2026-08-31 05:39:27 UTC…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-31T13:55:18+00:00).

• feat(rocm): add RDNA3 and RDNA4 runtime foundation: новый commit 45417560 → c0713e44.
• fix(rocm): make TVM-FFI index and store JIT kernels portable to HIP: новый commit 89ad7ab0 → 07a352a0.
• fix(rocm): gate CUDA-only optional backends on ROCm: новый commit 9a523dc8 → 32dfb7bd.
• fix(rocm): route tensor parallel communication through RCCL: новый commit 9d5e3fc8 → b794aa50.
• feat(rocm): enable native GGUF kernels on ROCm (validated on RDNA4): новый commit 49591978 → 7b3acf2f.
• FreeToken Roadmap (2026): body edited.
• Новый релевантный pr: feat(models): support TP for qwen3_5_moe.
• [Bug]: RuntimeError: Rowwise scaling is not currently supported on your device: обновлён (state open→open, comments 9→10).
• Новый релевантный pr: feat(qwen4_exp): stream the PLE n-gram table from disk.
• Новый релевантный issue: Windows/WDDM: offload serving fails when bank sources exceed the ~50%-RAM pageable-locking pool (repro, measurements, proposed path).
• Feature Request: Support Image Input: обновлён (state open→closed, comments 1→1).
• feat(qwen4-exp): mmap PLE tables from disc: обновлён (state open→open, comments 2→3).
• FlashML-org/FreeToken:main: 4b94bdc3 → 3a20a790 — fix(kernel): unbreak the nightly kernel-cache wheel build (fix(kernel): unbreak the nightly kernel-cache wheel build FlashML-org/FreeToken#310).
• Upstream main: релевантные изменения 4b94bdc3 → 3a20a790 — commits: fix(kernel): unbreak the nightly kernel-cache wheel build (fix(kernel): unbreak the nightly kernel-cache wheel build FlashML-org/FreeToken#310) | files: python/freetoken/moe/offload_cache.py.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aapache/tvm-ffi%20gfx1201&sort=updated&order=desc&per_page=20: HTTP 403: You have exceeded a secondary rate limit. Please wait a few minutes before you try again. For more on scraping GitHub and how it may affect your rights, please review our Terms of Service (https:…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-31T20:07:31+00:00).

• Новый релевантный pr: Fix GGUF user-defined token atomicity.
• ⚠️ 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 8402:325C1D:2E8B7B:301D7C:6A95DF27 and timestamp 2026-08-31 20:08:07 UTC. …

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-31T23:52:25+00:00).

• feat(models): support TP for qwen3_5_moe: обновлён (state open→open, comments 10→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 C423:2547C8:185277F:50A0CA5:6A9613D3 and timestamp 2026-08-31 23:52:51 UTC…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-09-01T05:12:52+00:00).

• Новый релевантный pr: qwen4_exp: load modelopt MIXED_PRECISION (NVFP4 experts + block-FP8 dense) checkpoints.
• Новый релевантный issue: Design discussion: first-class out-of-tree model providers.
• [Bug]: RuntimeError: Rowwise scaling is not currently supported on your device: обновлён (state open→open, comments 10→11).
• feat(qwen4_exp): stream the PLE n-gram table from disk: обновлён (state open→open, comments 1→2).
• Новый релевантный pr: fix(ftw): preserve auxiliary expert banks.
• Новый релевантный pr: fix(rocm): avoid indirect host pointers during HIP graph capture.
• FlashML-org/FreeToken:main: 3a20a790 → e05cff83 — perf(moe): route fused_topk through the in-repo triton router (perf(moe): route fused_topk through the in-repo triton router FlashML-org/FreeToken#319).
• Upstream main: релевантные изменения 3a20a790 → e05cff83 — commits: perf(moe): route fused_topk through the in-repo triton router (perf(moe): route fused_topk through the in-repo triton router FlashML-org/FreeToken#319) | files: python/freetoken/kernel/backend.py, python/freetoken/moe/fused.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 9431:23FB6:185A7C6:5061902:6A965EF4 and timestamp 2026-09-01 05:13:24 UTC.…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-09-01T11:41:30+00:00).

• feat(rocm): add RDNA3 and RDNA4 runtime foundation: обновлены описание/review/discussion (comments 3→4, review 0→0).
• Новый релевантный issue: Feature request: AMD GPU support.
• Новый релевантный issue: [BUG] qwen3_5 dense: mixed-precision NVFP4 crashes in ct_bf16_fuse (Float8 × BFloat16 promotion).
• Новый релевантный issue: Can't load Qwen3.8-Flash-Next.
• feat(qwen3_5_moe): support compressed-tensors NVFP4 experts with block-fp8 dense side: обновлён (state open→open, comments 0→1).
• fix(qwen3_5_moe): detect shared-expert quant for modelopt mixed checkpoints: обновлён (state open→open, comments 0→1).
• Новый релевантный issue: [Showcase & Fix] Qwen3.8-Flash-Next on Dual RTX 5090 Blackwell + Tool Call Recovery Patch.
• feat(kvcache): sub-byte Q4_0 and Q6_0 KV cache quantization: обновлён (state open→open, comments 3→4).
• ⚠️ 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 AC40:268019:3E1541B:CCBB3D7:6A96BA0D and timestamp 2026-09-01 11:42:05 UTC…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-09-01T16:41:47+00:00).

• Design discussion: first-class out-of-tree model providers: обновлён (state open→open, comments 1→1).
• fix(rocm): avoid indirect host pointers during HIP graph capture: обновлён (state open→open, comments 0→0).
• Новый релевантный pr: qwen3_5_moe: support unsloth/Qwen3.8-27B-NVFP4 (mixed-precision NVFP4+FP8).
• Новый релевантный pr: fix: [BUG] qwen3_5 dense: mixed-precision NVFP4 crashes in ct_bf16_fuse (Float8 × BFloat16 promotion).
• ⚠️ 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 71C1:1AFA92:9D033:20BE58:6A970077 and timestamp 2026-09-01 16:42:31 UTC. F…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-09-01T21:03:48+00:00).

• FreeToken Roadmap (2026): body edited.
• feat(qwen4_exp): stream the PLE n-gram table from disk: обновлён (state open→closed, comments 2→2).
• [Showcase & Fix] Qwen3.8-Flash-Next on Dual RTX 5090 Blackwell + Tool Call Recovery Patch: обновлён (state open→open, comments 1→2).
• Новый релевантный pr: feat(glm5_next): support GLM-5.3-Flash.
• feat(qwen4-exp): mmap PLE tables from disc: обновлён (state open→closed, comments 3→3).
• FlashML-org/FreeToken:main: e05cff83 → 4c0bad3f — feat(qwen4_exp): stream the PLE n-gram table from disk (feat(qwen4_exp): stream the PLE n-gram table from disk FlashML-org/FreeToken#311).
• Upstream main: релевантные изменения e05cff83 → 4c0bad3f — files: python/freetoken/engine/config.py, python/freetoken/engine/engine.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 5C05:3D375E:EAD31F:F28938:6A973DD7 and timestamp 2026-09-01 21:04:23 UTC. …

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-09-02T00:16:32+00:00).

• FreeToken Roadmap (2026): body edited.
• [Bug] Windows + RTX 50 (sm_120): RuntimeError: Rowwise scaling is not currently supported on your device during CUDA graph capture — dense FP8 W8A8 path: обновлён (state open→closed, comments 1→2).
• [Bug]: RuntimeError: Rowwise scaling is not currently supported on your device: обновлён (state open→open, comments 11→12).
• FreeToken-Intel: 35B MoE, live tokens off an Arc Pro B70 (SYCL / XPU port): обновлён (state open→open, comments 0→0).
• Feature request: AMD GPU support: обновлён (state open→open, comments 15→17).
• Proposal: reproducible end-to-end benchmark / regression harness: обновлён (state open→open, comments 3→4).
• 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).
• feat(glm5_next): support GLM-5.3-Flash: обновлён (state open→closed, comments 0→0).
• FlashML-org/FreeToken:main: 4c0bad3f → a2538a42 — feat(glm5_next): support GLM-5.3-Flash (feat(glm5_next): support GLM-5.3-Flash FlashML-org/FreeToken#332).
• Upstream main: релевантные изменения 4c0bad3f → a2538a42 — files: python/freetoken/engine/engine.py, python/freetoken/kernel/triton/activation.py, python/freetoken/kernel/triton/glm_dsa_sparse.py, python/freetoken/kernel/triton/kpool_compress.py, python/freetoken/kernel/triton/mhc.py, python/freetoken/kvcache/init.py, python/freetoken/kvcache/dsa_pool.py, python/freetoken/moe/benchbw.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 1447:B978C:36F3700:B34BFB7:6A976B09 and timestamp 2026-09-02 00:17:13 UTC.…

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

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-09-02T04:45:03+00:00).

• fix(download): include json files when fetching weights, so NVFP4's index.json is no longer silently skipped: state open → closed.
• fix(download): include json files when fetching weights, so NVFP4's index.json is no longer silently skipped: обновлены описание/review/discussion (comments 2→3, review 0→0).
• NVFP4 checkpoints fail to load: model.safetensors.index.json is not downloaded, but the NVFP4 bank loader requires it: state open→closed.
• Новый релевантный pr: feat(rocm): AMD ROCm support (gfx1100) + Qwen3.5-MoE GGUF loader.
• Новый релевантный pr: fix(engine): reserve the pool's dummy page in the --moe-cache-auto KV floor.
• Новый релевантный pr: fix(kernels): drop the D2H sync from the varlen GDN/KDA prefill conv.
• Новый релевантный pr: perf(ple): fuse the n-gram row-id hash into one Triton kernel.
• Proposal: reproducible end-to-end benchmark / regression harness: обновлён (state open→open, comments 4→7).
• Новый релевантный pr: feat(moe): NVMe disk tier for MoE expert banks.
• fix(models): load NVFP4 expert banks from checkpoints without a safetensors index: обновлён (state open→closed, comments 0→1).
• feat(models): support TP for qwen3_5_moe: обновлён (state open→open, comments 11→12).
• FlashML-org/FreeToken:main: a2538a42 → a80b4d30 — fix(hf): download the shards the safetensors index names (fix(hf): download the shards the safetensors index names FlashML-org/FreeToken#336).
• ⚠️ 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 5400:168969:2614347:276478E:6A97A9F1 and timestamp 2026-09-02 04:45:37 UTC…

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