| commit |
1.00 |
Fix MoE expert decompression for non-32-divisible bit widths |
|
Commit message contains explicit AI assi |
2026-07-22 |
| commit |
1.00 |
Fix double-shifted training loss in GitForCausalLM (#47395) |
|
Commit message contains explicit AI assi |
2026-07-21 |
| commit |
1.00 |
Fix `TrackioCallback` fails to log evaluation metrics after |
|
Commit message contains explicit AI assi |
2026-07-21 |
| commit |
1.00 |
Fix orthogonal_ init for low-precision dtypes (bf16/fp16) (# |
|
Commit message contains explicit AI assi |
2026-07-17 |
| commit |
1.00 |
tests: reduce processor test memory usage by using tiny Hub |
|
Commit message contains explicit AI assi |
2026-07-10 |
| commit |
1.00 |
Fix flash-attn Docker build broken by setuptools 83 removing |
|
Commit message contains explicit AI assi |
2026-07-10 |
| pr |
1.00 |
Fix CodeLlama tokenizer dropping leading whitespace on decod |
|
PR body explicitly mentions AI collabora |
2026-07-22 |
| pr |
1.00 |
Exportable kimi |
|
PR body explicitly mentions AI collabora |
2026-07-06 |
| pr |
1.00 |
[Mistral] Add tekken tokenizer detection, conversion, and sa |
|
PR body explicitly mentions AI collabora |
2026-06-12 |
| pr |
1.00 |
adding amd quark config class changes |
|
PR body explicitly mentions AI collabora |
2026-07-14 |
| pr |
1.00 |
docs: add overflow_to_sample_mapping to tokenizer return val |
|
PR body explicitly mentions AI collabora |
2026-07-22 |
| pr |
1.00 |
Kernels and loaders robustification |
|
PR body explicitly mentions AI collabora |
2026-07-15 |
| pr |
1.00 |
pytorch 2.13 is released , updating transformer to handle th |
|
PR body explicitly mentions AI collabora |
2026-07-09 |
| pr |
1.00 |
CPU can incur a slow path on non-contiguous magnitudes |
|
PR body explicitly mentions AI collabora |
2026-07-15 |
| pr |
1.00 |
[Quantization]: Refactor is_quantization_compressed for form |
|
PR body explicitly mentions AI collabora |
2026-07-08 |
| pr |
1.00 |
fix: liger unnecessarily materializes logits in VRAM during |
|
PR body explicitly mentions AI collabora |
2026-04-06 |
| pr |
1.00 |
Executorch exporter fixes |
|
PR body explicitly mentions AI collabora |
2026-07-10 |
| pr |
1.00 |
Fix causality violation in Zamba2 inter-chunk state recurren |
|
PR body explicitly mentions AI collabora |
2026-07-22 |
| pr |
1.00 |
Fix finegrained-fp8 offline loading bug |
|
PR body explicitly mentions AI collabora |
2026-06-04 |
| pr |
1.00 |
Batch Rebalance Data Sampler |
|
PR body explicitly mentions AI collabora |
2026-07-15 |
| pr |
1.00 |
[Auto-Fix] Fix failing tests for inkling |
|
PR body explicitly mentions AI collabora |
2026-07-21 |
| pr |
0.20 |
Lag kv cache |
|
Slightly more explanation, but domain-sp |
2025-05-26 |
| commit |
0.10 |
CPU can incur a slow path on non-contiguous magnitudes (#473 |
|
Technical, domain-specific, concise; no |
2026-07-22 |
| commit |
0.10 |
Route chat management calls to the service root (#47138) (#4 |
|
Informative, detailed, with technical ja |
2026-07-22 |
| commit |
0.10 |
Fix gradient inflation when combining label smoothing with g |
|
Analysis is detailed and technical with |
2026-07-22 |
| commit |
0.10 |
[Qwen3ASR] Add hotword parsing, and fix language parsing and |
|
Domain-specific, with informal amendment |
2026-07-22 |
| commit |
0.10 |
Fix multi-device mxfp4 dequantization race in `_convert_moe_ |
|
Technical, signed, multi-author, and dom |
2026-07-22 |
| commit |
0.10 |
Fix CohereASR training-loss double-shift (same as Moonshine |
|
Informal yet technical commit style; spe |
2026-07-21 |
| commit |
0.10 |
Add distributed runtime utils and DistributedMixin (#47352) |
|
Commit history with terse notes suggests |
2026-07-21 |
| commit |
0.10 |
Fix inkling feature extractor (#47349) |
|
Uses domain jargon and informal edits, n |
2026-07-15 |
| commit |
0.10 |
ci: cover xet as well (runtime error) (#47338) |
|
Technical and concise with minor typos, |
2026-07-15 |
| commit |
0.10 |
Pin kernels to compatible version in benchmark workflow (#47 |
|
Technical, detailed changelog, consisten |
2026-07-15 |
| pr |
0.10 |
Fix loss alignment and Trainer token counting for encoder de |
|
Title is specific and technical; content |
2026-06-26 |
| pr |
0.10 |
Better and more extensive tests for RoPE |
|
Brief, domain-specific title; no verbose |
2026-06-26 |
| pr |
0.10 |
🚨 Enable SDPA (and other attention backends) for T5 and prop |
|
Short, emoji-enhanced technical title; c |
2026-07-02 |
| pr |
0.10 |
:rotating_light: Processors update the rest |
|
Title uses emoji and short phrase; no AI |
2026-06-11 |
| pr |
0.10 |
Clarify device placement in pipelines |
|
Concise, technical title; no AI-generati |
2026-07-16 |
| pr |
0.10 |
[Offloading] [Bugfix] Fix fully offloaded model saving |
|
Bracketed, bugfix-style title; typical h |
2026-07-15 |
| pr |
0.10 |
Fix broken relative links in benchmark and i18n READMEs |
|
Technical, concise change log; process n |
2026-07-22 |
| pr |
0.10 |
[generation] Encode multimodal data only once |
|
Short, domain-tagged title; no verbose o |
2026-05-05 |
| pr |
0.10 |
Add V-JEPA 2.1 inference support |
|
Technical, inference-specific title; hum |
2026-04-17 |
| pr |
0.10 |
feat: update GGUF model configuration parameters for Gemma 4 |
|
Brief technical note, some formality, bu |
2026-06-06 |
| pr |
0.10 |
fix: allow tokenizers 0.23.1 to unblock security fix |
|
Brief, technical motif with minimal moti |
2026-07-21 |
| commit |
0.05 |
[Quantization]: Refactor is_quantization_compressed for form |
|
Standard commit format, technical and co |
2026-07-22 |
| commit |
0.05 |
CI: add serge verify (GPU) caller workflow (#47381) |
|
Highly technical, terse, natural domain |
2026-07-17 |
| commit |
0.05 |
ci: harden daily CI reporting against GitHub API rate limits |
|
Informal, domain-specific, includes erro |
2026-07-17 |
| pr |
0.05 |
[Model] Add LingBot-Vision Model |
|
Slightly more elaborate, but concise and |
2026-07-16 |
| commit |
0.00 |
[fix] fix requirements audio feature and proc (#47113) |
|
Very terse, telegraphic, and domain-like |
2026-07-22 |
| commit |
0.00 |
fix: liger unnecessarily materializes logits in VRAM during |
|
Single line, domain-specific, terse; no |
2026-07-22 |
| commit |
0.00 |
Remove deprecated training args and `is_fast` property (#469 |
|
Terse, domain-standard, includes sign-of |
2026-07-22 |
| commit |
0.00 |
Consistent output shape from `get_image_features` (#46405) |
|
One-word lines and shorthand, highly inf |
2026-07-22 |
| commit |
0.00 |
fix failed test cases for qwen3_omni_moe model (#47449) |
|
Commit messages are brief, specific, and |
2026-07-22 |
| commit |
0.00 |
Fix Hunyuan-VL PIL image resize parity with reference prepro |
|
Very terse human-like commit messages wi |
2026-07-22 |
| commit |
0.00 |
Move `value` padding into the attention interfaces that need |
|
Human-typical, terse, domain-specific co |
2026-07-22 |
| commit |
0.00 |
Simplify function dispatch for linear attention (#47450) |
|
Short, informal, domain-specific commit |
2026-07-22 |
| commit |
0.00 |
[cache] Allow sliding window layers to be roll-backed for sp |
|
Informal, specific, and human-style comm |
2026-07-22 |
| commit |
0.00 |
Warn when `group_by_length` is silently ignored for iterable |
|
Concise, terse commit message typical of |
2026-07-21 |
| commit |
0.00 |
Update bug report list (#46607) |
|
Short, human-like commit message and che |
2026-07-21 |
| commit |
0.00 |
Fix shape mismatch in KyutaiSpeechToText `generate()` last w |
|
Terse commit messages and use of signed- |
2026-07-21 |
| commit |
0.00 |
Optimize flash attention max seqlen computation in vision at |
|
Terse changelog messages with human-like |
2026-07-21 |
| commit |
0.00 |
fix: remove unreachable return in special token builder (#47 |
|
Minimal, clear, informal message as is c |
2026-07-21 |
| commit |
0.00 |
Add Harry to slow CI (#47454) |
|
Single, brief, informal comment typical |
2026-07-21 |
| commit |
0.00 |
BLT: vectorize patch length processing (#47385) |
|
Changelog-style terse message with domai |
2026-07-21 |
| commit |
0.00 |
[Kimi] add integration tests (#47383) |
|
Terse style with non-standard capitaliza |
2026-07-21 |
| commit |
0.00 |
Fix typo in `MusicgenForCausalLM.generate()` (#46974) |
|
Minimal, concise changelog entries with |
2026-07-21 |
| commit |
0.00 |
Fix Cosmos 3 Edge Patch packing order (#47399) |
|
Commit messages are terse, technical, an |
2026-07-20 |
| commit |
0.00 |
Fix yarn `mscale_all_dim` for DeepSeek v2 and Mistral 4 (#47 |
|
Commit messages are concise, domain-spec |
2026-07-20 |
| commit |
0.00 |
Fix recurrent-layer padding mask being skipped on continued |
|
Technical explanations with natural edit |
2026-07-20 |
| commit |
0.00 |
Hoist special-token lookups out of per-token loops in six sl |
|
Succinct, domain-jargon usage, no signal |
2026-07-20 |
| commit |
0.00 |
fix typos and variable naming in quicktour.md (#47418) |
|
Typical human edit messages and casual l |
2026-07-20 |
| commit |
0.00 |
Normalize multimodal input keys in AnyToAnyPipeline (#47074) |
|
Domain-specific, includes typical commit |
2026-07-20 |
| commit |
0.00 |
Fix Aria checkpoint key conversion mapping (#47151) |
|
Detailed technical note, domain-specific |
2026-07-20 |
| commit |
0.00 |
extend tests/models/qwen3_next/test_modeling_qwen3_next.py:: |
|
Commit is technical, includes email, no |
2026-07-20 |
| commit |
0.00 |
Fix `BlockMask` crash in CTRL flex-attention generation (#46 |
|
Extremely terse, cryptic commit message; |
2026-07-17 |
| commit |
0.00 |
[docs] Fix decode examples and expected output in fast_token |
|
Minimal, typical doc fix, no sign of AI |
2026-07-17 |
| commit |
0.00 |
Fix image-text-to-text stop_sequence handling (#47032) |
|
Standard fix, includes human co-author, |
2026-07-17 |
| commit |
0.00 |
[serge] Fix 20 integration tests for model `whisper` failing |
|
— |
2026-07-17 |
| commit |
0.00 |
[Mistral] Move MistralConverter into integrations/mistral/ p |
|
Regular refactor commit, normal structur |
2026-07-17 |
| commit |
0.00 |
Fix GLM video frame padding for temporal patches (#47141) |
|
Technical changelog, natural tone, speci |
2026-07-17 |
| commit |
0.00 |
Refactor image loading in tests to use load_test_image helpe |
|
Human-like granular commit list, non-AI |
2026-07-17 |
| commit |
0.00 |
Add Cosmos3 Edge model support (#47181) |
|
Detailed technical changelog, natural it |
2026-07-16 |
| commit |
0.00 |
[docs] Inkling (#47350) |
|
Minimal, ambiguous message, matches huma |
2026-07-16 |
| commit |
0.00 |
Fix Daily CI reporting issues (#47364) |
|
Concise, domain-specific language; no AI |
2026-07-16 |
| commit |
0.00 |
[AudioFlamingo3] Fix bfloat16 dtype mismatch in audio encode |
|
Detailed technical description with jarg |
2026-07-16 |
| commit |
0.00 |
[`peft`] Support key_mapping with PEFT models (#46766) |
|
Domain-specific, co-authored, natural hu |
2026-07-16 |
| commit |
0.00 |
Fix model tests for tipsv2 (#47356) |
|
Human tone, specific file/test reference |
2026-07-16 |
| commit |
0.00 |
Fix TimesFM 2.5 window_size AttributeError (#47363) |
|
Terse and informal; focus on bug and tes |
2026-07-16 |
| commit |
0.00 |
fix: allow num_labels property to return None when id2label |
|
Domain detail, signed-off-by trailer, co |
2026-07-16 |
| commit |
0.00 |
[Tests] Fix slow video tensor creation from list of numpy ar |
|
Standard test fix wording, informal, dom |
2026-07-16 |
| commit |
0.00 |
Update dev version on main (#47366) |
|
Single-word summary, typical for human v |
2026-07-16 |
| commit |
0.00 |
[CB] Automatically switch attention implementation to flash |
|
Human commit log style, bullet points, d |
2026-07-16 |
| commit |
0.00 |
Fix deepgemm on multiple devices (#47323) |
|
Colloquial ('claude's fix'), informal, d |
2026-07-16 |
| commit |
0.00 |
[`FP8`] Bump kernels version (#47344) |
|
Terse one-word commit message; human sty |
2026-07-16 |
| commit |
0.00 |
ci: Handle empty GitHub token in CI run lookup (#47362) |
|
Uses typical developer commit phrasing; |
2026-07-16 |
| commit |
0.00 |
Fix assisted decoding for models with EncoderDecoder cache & |
|
Message is terse, informal, and includes |
2026-07-16 |
| commit |
0.00 |
Fix sdpa prefill with position_bias (#47359) |
|
Brief message with technical terms; not |
2026-07-16 |
| commit |
0.00 |
[docs] FlashAttention kernel fallback (#47345) |
|
Single-word commit messages, no AI indic |
2026-07-15 |
| commit |
0.00 |
Small improvement (#47348) |
|
Extremely terse edit, clearly human-writ |
2026-07-15 |
| commit |
0.00 |
Add Inkling model (#47347) |
|
Highly informal, with typos and in-progr |
2026-07-15 |
| commit |
0.00 |
🚨 Pass tp_plan from lm_heads (#47253) |
|
Colloquial tone and test-related errors |
2026-07-15 |
| commit |
0.00 |
[docs] TokenizersBackend fallback (#47302) |
|
Template signal; no free text to analyze |
2026-07-15 |
| commit |
0.00 |
Fix models for which we don't have a dedicated tokenizer cla |
|
Very informal, with typos and colloquial |
2026-07-15 |
| commit |
0.00 |
Resolve continuous batching XPU availability checks at runti |
|
Standard developer commits, no free-text |
2026-07-15 |
| commit |
0.00 |
[Nit] Add kernels_fallback_ok kwarg to is_flash_attn_N_avail |
|
Short commit titles, domain nits, and te |
2026-07-15 |
| commit |
0.00 |
[Nit] Add expectations for gemma4 tests on H100 (#47311) |
|
Terse, domain-specific commit titles; no |
2026-07-15 |
| commit |
0.00 |
[docs] DeepGEMM requirements (#47324) |
|
Commit message is terse, references doma |
2026-07-14 |
| commit |
0.00 |
:rotating_light: Fix GPTBigCode and GPTNeoX for the Transfor |
|
Human-like informal tone, includes domai |
2026-07-14 |
| commit |
0.00 |
DeepGEMM shouldn't pad on SM90 (#47313) |
|
The brief 'claude's fix' is highly infor |
2026-07-14 |
| commit |
0.00 |
Fix half-precision torch.compile crash in DETR-family sine p |
|
Detailed technical explanation and chang |
2026-07-14 |
| commit |
0.00 |
Fix hardcoded paths in siglip checkpoint/vocab loading (#471 |
|
Terse, issue-focused commit message, no |
2026-07-14 |
| commit |
0.00 |
[generation] Fix misleading synced_gpus warning in continuou |
|
Describes a fix and test update with hum |
2026-07-14 |
| commit |
0.00 |
Update AMD CI runner groups to amd-mi300 (#47307) |
|
Short and specific update message; no AI |
2026-07-14 |
| commit |
0.00 |
Point to Gemma 4 model in Gemma4ForCausalLM docstring exampl |
|
Docstring update noted concisely, inform |
2026-07-13 |
| commit |
0.00 |
Fix Qwen Omni batched text postprocessing (#47197) |
|
Technical domain language and brief sign |
2026-07-13 |
| commit |
0.00 |
Fix AqlmConfig error messages to say "int" instead of "float |
|
Problem-focused commit, uses informal ex |
2026-07-13 |
| commit |
0.00 |
Fix check for interactive stdout in _style function (#47283) |
|
Concise, domain-specific fix summary in |
2026-07-13 |
| commit |
0.00 |
Fix get_json_schema crash on non-string docstring choices (# |
|
Technical explanation, domain abbreviati |
2026-07-13 |
| commit |
0.00 |
Make `MODEL_IDS_TO_TOKENIZERS_BACKEND` capture all DeepSeek |
|
Commit trailer, standard style, terse la |
2026-07-13 |
| commit |
0.00 |
Update doc preprocessing regex to prevent ReDoS (#47187) |
|
Short commit description, no AI signals, |
2026-07-13 |
| commit |
0.00 |
[generate] Add proper MTP support (#46229) |
|
Terse, informal progress logs, no AI pat |
2026-07-13 |
| commit |
0.00 |
Shard on read Dtensor aware (#46717) |
|
Contains typos, domain jargon, informal |
2026-07-13 |
| commit |
0.00 |
Switch AMD daily CI to mi300 runners (#47259) |
|
Terse and technical commit message; huma |
2026-07-11 |
| commit |
0.00 |
Torch compile backend defaults to "neuron" (#47035) |
|
Commit messages are brief, use domain te |
2026-07-10 |
| commit |
0.00 |
Add heterogeneous config support (per-layer configuration) ( |
|
Commit uses domain jargon, succinct huma |
2026-07-10 |
| commit |
0.00 |
[fix] update integration test values (#47146) |
|
Minimal, terse commit language; nothing |
2026-07-10 |
| commit |
0.00 |
Fix DeepSpeed SP loss aggregation and LocalLayerRepository k |
|
Detailed but technical, with typos and i |
2026-07-10 |
| commit |
0.00 |
Fix FA performance regression (#47134) |
|
Brief notes, domain language, co-author |
2026-07-10 |
| commit |
0.00 |
tests only for the top 10 download models (#47244) |
|
Very terse, human style; no AI hallmarks |
2026-07-10 |
| pr |
0.00 |
fix: guard DTensor import for PyTorch < 2.5 compatibility |
|
Only template shown; no AI signals in th |
2026-07-22 |
| pr |
0.00 |
Port ESMC and ESMFold2 to Transformers |
|
Only template shown; PR content is just |
2026-06-04 |
| pr |
0.00 |
Add Granite-swa and Granitemoe-swa model support |
|
Template plus brief, domain-typical titl |
2026-07-08 |
| pr |
0.00 |
Add Fun-ASR-Nano model |
|
Template plus only a short model-related |
2026-05-24 |
| pr |
0.00 |
[Bugfix] [Modeling] Fix DSV4 norm |
|
Just a bugfix title and template; no AI- |
2026-07-22 |
| pr |
0.00 |
Fix Qwen2.5-Omni Token2Wav DiT rotary embedding layout (inte |
|
Brief technical title; remaining is temp |
2026-07-18 |
| pr |
0.00 |
wip step 3 7 flash |
|
Clear domain-specific project notes, inf |
2026-06-15 |
| pr |
0.00 |
Implement VibeVoice |
|
Just a project name in the title; no con |
2025-08-29 |
| pr |
0.00 |
Isolate MLA KV expansion to make it easier to bypass |
|
Very concise, domain-specific, no AI sty |
2026-07-21 |
| pr |
0.00 |
Support Granite Speech NAR (NLE) |
|
Only a PR title, no free-text content to |
2026-05-18 |
| pr |
0.00 |
add MTP for GLM OCR and other MROPEs |
|
Only a title is provided; no AI signals |
2026-07-21 |
| pr |
0.00 |
add HyperClovaX Vision |
|
Simple, lowercase title; consistent with |
2026-02-27 |
| pr |
0.00 |
[WIP] PrismML Bonsai model support |
|
Only a PR title, no free-text descriptio |
2026-04-01 |
| pr |
0.00 |
[fix] fix requirements audio feature and proc |
|
Only a PR title filled, no AI signals pr |
2026-07-06 |
| pr |
0.00 |
Update serve chat parsing |
|
Only a title, without any free-text cont |
2026-05-28 |
| pr |
0.00 |
Also size the device_map buffer from the largest leaf module |
|
Title only, no signs of AI-generated tex |
2026-07-09 |
| pr |
0.00 |
model: Add NVIDIA Canary-1B-v2 to Transformers |
|
Only title provided; no AI traits detect |
2026-06-22 |
| pr |
0.00 |
[fix] inkling: mps + cuda mel spec extraction |
|
No free-text, terse technical title; hum |
2026-07-20 |
| pr |
0.00 |
add_axk1 |
|
Very terse, technical title only; not AI |
2026-06-24 |
| pr |
0.00 |
Route chat management calls to the service root (#47138) |
|
Just a title, no free-form AI-like text. |
2026-07-13 |
| pr |
0.00 |
Remove `is_custom_code` gate from per-param `_is_hf_initiali |
|
Only a brief PR title with no free-text |
2026-07-15 |
| pr |
0.00 |
fix: guard DTensor import in sharding_utils.py for PyTorch < |
|
Only includes a brief, domain-specific P |
2026-07-22 |
| pr |
0.00 |
feat[vLLM x v5]: Make audio optional and support multiple au |
|
PR title uses domain-specific abbreviati |
2026-07-22 |
| pr |
0.00 |
LocateAnything 3B |
|
Short title, terse style, no AI hallmark |
2026-05-30 |
| pr |
0.00 |
Fix Gemma4 audio feature dtype mismatch in masked_scatter |
|
Title uses detailed domain terms, no AI |
2026-07-22 |
| pr |
0.00 |
Fix gradient inflation when combining label smoothing with g |
|
A short, technical PR title with no AI h |
2026-07-11 |
| pr |
0.00 |
Fix MoE expert decompression for non-32-divisible bit widths |
|
Only a technical fix title, no AI-typica |
2026-07-14 |
| pr |
0.00 |
[Qwen3ASR] Add hotword parsing, and fix language parsing and |
|
Technical, concise title; no AI text pre |
2026-07-06 |
| pr |
0.00 |
Remove deprecated training args and `is_fast` property |
|
Only a domain-specific, succinct PR titl |
2026-06-26 |
| pr |
0.00 |
Add MOSS-TTS-v1.5 (delay) and MOSS-Audio-Tokenizer |
|
PR title only, technical, no AI-generate |
2026-06-05 |
| pr |
0.00 |
Fix CLVP `c_proj` residual scaled-init silently never applie |
|
Uses domain jargon and is specific, non- |
2026-07-22 |
| pr |
0.00 |
CI: fix torchaudio pinning +proper break in rnnt |
|
Domain-specific, brief, informal structu |
2026-07-20 |
| pr |
0.00 |
Don't mutate class-level common_kwargs defaults in _merge_kw |
|
Technical and concise, no free-text sugg |
2026-07-22 |
| pr |
0.00 |
Fix stale phonemizer_lang in Wav2Vec2PhonemeCTCTokenizer |
|
Only a terse, domain-specific title with |
2026-07-21 |
| pr |
0.00 |
fix failed test cases for qwen3_omni_moe model |
|
Only title present; no AI signals in bri |
2026-07-21 |
| pr |
0.00 |
Align all mambas to latest best standards for much more cove |
|
Title only; uses casual, domain language |
2026-07-21 |
| pr |
0.00 |
Re-enable VITS model forward test |
|
Short, direct title, typical of human co |
2026-07-06 |
| pr |
0.00 |
Fix RT-DETR v2 converter for current model keys |
|
Brief, technical phrasing in title only; |
2026-07-07 |
| pr |
0.00 |
Add LocateAnything |
|
Contains domain-specific terms and infor |
2026-06-19 |
| pr |
0.00 |
Add Molmo2 |
|
Only a brief title, no free-text to eval |
2026-01-23 |
| pr |
0.00 |
[serge] Fix 28 integration tests for model `whisper` failing |
|
— |
2026-07-22 |
| pr |
0.00 |
CI: use a single function for GH calls |
|
No free-text content, just title with te |
2026-07-22 |
| pr |
0.00 |
Consistent output shape from `get_image_features` |
|
Only a technical title, no verbose or AI |
2026-06-04 |
| pr |
0.00 |
Fix multi-device mxfp4 dequantization race in `_convert_moe_ |
|
The PR title is terse and domain-specifi |
2026-07-20 |
| pr |
0.00 |
extend tests/quantization/gptq/test_gptq.py::GPTQTestCUDA an |
|
PR title is descriptive, lacks AI hallma |
2026-07-08 |
| pr |
0.00 |
Fix Hunyuan-VL PIL image resize parity with reference prepro |
|
Short, domain-specific title; template o |
2026-07-10 |
| pr |
0.00 |
Fix ESMFold error when explicitly cast to fp16 |
|
Title only, matches typical terse engine |
2026-07-22 |
| pr |
0.00 |
Use new `per_layer_config` for Gemma 4 so that heterogeneous |
|
Title only, highly technical; no AI sign |
2026-07-17 |
| pr |
0.00 |
Fix ESMFold error when explicitly cast to fp16 |
|
Same as previous: very brief, direct fix |
2026-07-22 |
| pr |
0.00 |
Move `value` padding into the attention interfaces that need |
|
Compact, technical, no markers of AI out |
2026-07-21 |
| pr |
0.00 |
Deprecate CB config in gen configuration |
|
Only title; clear, concise technical lan |
2026-07-13 |
| pr |
0.00 |
Fix torch_float casting to int instead of float |
|
Title is terse and domain-specific with |
2026-07-22 |
| pr |
0.00 |
Simplify function dispatch for linear attention |
|
Contains just a terse, domain-specific t |
2026-07-21 |
| pr |
0.00 |
[cache] Allow sliding window layers to be roll-backed for sp |
|
Technical, project-specific title; human |
2026-07-21 |
| pr |
0.00 |
Add FSDP plans to all models |
|
Technical changelog, direct and domain-s |
2026-07-08 |
| pr |
0.00 |
Add FSDP CI and end-to-end FSDP tests + save fsdp |
|
Only a brief, technical title; no stylis |
2026-07-16 |
| pr |
0.00 |
[serge] Fix 12 integration tests for model `glm46v` failing |
|
— |
2026-07-13 |
| pr |
0.00 |
Add MXFP4 training support via a dequantize-on-the-fly fallb |
|
Title shows domain jargon and lacks AI p |
2026-07-21 |
| pr |
0.00 |
. |
|
Single '.' as title is minimal and plain |
2026-07-21 |
| pr |
0.00 |
Fix double-shifted training loss in GitForCausalLM |
|
Only a technical title; no AI-typical la |
2026-07-18 |
| pr |
0.00 |
Add chat input sanitization |
|
Short and technical; lacks AI-generated |
2026-07-17 |
| pr |
0.00 |
Add support for NeuCodec |
|
Concise with technical reference, not AI |
2026-07-07 |
| pr |
0.00 |
[DiffusionGemma] Support gradient checkpointing |
|
Domain shorthand in brackets; tone is hu |
2026-06-11 |
| pr |
0.00 |
[DiffusionGemma] Cast the decoder padding mask to bool |
|
Terse, domain-specific with natural shor |
2026-07-13 |
| pr |
0.00 |
[DeepSpeed] Composite models are not initialized correctly u |
|
Only a terse, domain-specific PR title w |
2026-07-13 |
| pr |
0.00 |
Fix GPT-2 c_proj depth scaling initialization |
|
Contains only a succinct technical PR ti |
2026-07-21 |
| pr |
0.00 |
Deprecate the old response_schema |
|
Simple deprecation-focused title, zero A |
2026-07-14 |
| pr |
0.00 |
Fix CohereASR training-loss double-shift (same as Moonshine |
|
Concise reference style, links to previo |
2026-06-25 |
| pr |
0.00 |
[docs] MPS graph cache |
|
Short, domain-specific docs note; no evi |
2026-07-13 |
| pr |
0.00 |
Warn when `group_by_length` is silently ignored for iterable |
|
Brief warning-focused title without Chat |
2026-07-17 |
| pr |
0.00 |
fix: correct typos in comments |
|
Only a terse changelog-style title and n |
2026-07-21 |
| pr |
0.00 |
Fix Whisper integration tests dtype mismatch and max_length |
|
Only PR template and human-written title |
2026-07-20 |
