diff --git "a/index.html" "b/index.html" --- "a/index.html" +++ "b/index.html" @@ -1,9 +1,9 @@ transformers radar

transformers downstream radar ยท 2026-11-02

44 repos ยท last 2 days of GitHub activity ยท 662 monkey patches tracked

Act on these

-
actioncomponentwhatthreads
๐Ÿ”ดfix in transformersmodel:GemmaGemma 4 FP8 quantization corrupts image input due to premature FP8 casting in patch_embeddervllm-project/vllm#63732
๐Ÿ”ดfix in transformersmodel:DeepSeek-V4DeepSeek-V4-Pro MoE produces garbled output under tensor parallelism due to incorrect expert routingvllm-project/vllm#47528
๐Ÿ”ดfix in transformerstokenizersGGUF loading fails for models with non-dict extra_special_tokens in tokenizer_config.jsonggml-org/llama.cpp#31343
๐Ÿ”ดfix in transformersmodel:GLMGLM-5.3-Flash fails under pipeline parallelism due to stale top-k indices in reused layersvllm-project/vllm#63170
๐Ÿ”ดfix in transformersmodel:QwenQwen3-VL vision features diverge from Transformers due to Python-dependent grid/token type handlingsgl-project/sglang#35772 huggingface/lerobot#5163
๐Ÿ”ดreply/guide usersfrom_pretrainedFAST action tokenizers from transformers 4 fail to load in transformers 5 due to backend instantiation changeshuggingface/lerobot#5155
๐ŸŸ fix in transformersquantizersBitsAndBytes 4-bit unmerge loses scale data causing progressive weight drifthuggingface/peft#4069
๐ŸŸ fix in transformersmodel:GLMGLM-5.x models fail to load from GGUF due to missing quant_config in embedding layervllm-project/vllm#37220
๐ŸŸ fix in transformerstokenizersTokenizerManager blocks async event loop during long fast-tokenizer callssgl-project/sglang#30771
๐ŸŸ reply/guide usersfrom_pretrainedGR00T fine-tunes require base model download even when weights are overwrittenhuggingface/lerobot#4778
๐ŸŸ fix in transformersmodel:GemmaGemma4Plugin audio processing fails due to inconsistent feature_extractor handlinghiyouga/LlamaFactory#10388
๐ŸŸ reply/guide usersfrom_pretrainedLoRA target_modules patterns use re.fullmatch which is not obvious to usershuggingface/peft#4067
๐ŸŸ fix in transformersmodel:QwenQwen-VL video loaders ignore explicit num_frames parametervllm-project/vllm#64899 vllm-project/vllm#64882
โšชfix in transformersmodel:QwenQwen2-VL rejects empty videos causing division by zero in frame samplingvllm-project/vllm#58379
โšชreply/guide userstokenizersPeftModelForQuestionAnswering silently drops token_type_ids and position_idshuggingface/peft#3857
-
All 321 matching threads -
threadcommentstitle
huggingface/peft#2310174Comparison of Different Fine-Tuning Techniques for Conversational AI
vllm-project/vllm#50576124[Feature]: SM8x (Ampere A100/A800) support for DeepSeek-V4-Flash / DeepSeek-V4-Flash-0731 DSpark
unslothai/unsloth#1473040Studio: Flows, a node editor for AI pipelines, story games and 2D worlds
vllm-project/vllm#5329830[Bugfix][KV Connector] Restore state after KV load rejection
huggingface/lerobot#500830feat(policies): add native Long-WAM training and inference
sgl-project/sglang#4609127[Feat] usage: align prompt and cache accounting across APIs
sgl-project/sglang#4191625[Fix] sampling: Make top-k/top-p filtering consistent across sampling paths
vllm-project/vllm#6370021[Bugfix][MoE] Avoid double router weighting in TRTLLM BF16 experts
vllm-project/vllm#5615520[Core][Attention] Resolve per-layer KV dtype in attention runtime
sgl-project/sglang#4722919[Fix] kernels: Support OOT fused ops under torch.compile
unslothai/unsloth#1607118LoRA: add the rank-r term inside the base GEMM so small updates are not rounded away
verl-project/verl#748317[vllm] feat: Support LoRA for DeepSeek V4 & DeepSeek V4.1 & GLM5.3 Flash
vllm-project/vllm#5594716[Feature][Spec Decode] Add Uno shared-model drafting with a draft-only LoRA adapter (Model Runner V2)
vllm-project/vllm#5731113[Perf][Startup] Move atomic_writer out of weight_utils so the model-info cache save stays light
vllm-project/vllm#3185613[Bug]: Text degeneration / repetition loops with MiniMax-M2.1-NVFP4 on v0.14.0rc1.dev308
sgl-project/sglang#4685813[Feat] lfm3: Add text, vision, audio and Omni support
sgl-project/sglang#3605913[Perf] dsv4: Add native H16 CuTe sparse prefill backend
sgl-project/sglang#3077113[Fix] tokenizer: Isolate and offload regular fast tokenization
unslothai/unsloth#1603212Opt-in fused batch-1 NF4 decode for Llama (UNSLOTH_DECODE_FUSED_NF4=1): 1.6-2.3x per token in eager generate
unslothai/unsloth#1622712Studio MLX: decode reply snapshots incrementally instead of re-decoding every id
vllm-project/vllm#5969311[ROCm][Perf] Kimi-K3: token-sharded residual stream for long prefills
unslothai/unsloth#711511Add EXL3 (ExLlamaV3) quantization backend: 2/3/4/6/8-bit and MoE support
sgl-project/sglang#4562111[Feat] decisions: Serve StartLux-Decision checkpoints on /v1/systemone
sgl-project/sglang#4750311[Perf] rmsnorm: One SM100-tuned JIT template for plain, fused residual-add and fp8 per-tensor quant
sgl-project/sglang#3322611[Feat] cp: Support FlashInfer prefill context parallelism
sgl-project/sglang#4488311[Feat] [NPU] multimodal: Enable device-side multimodal feature transport via torch_npu IPC
verl-project/verl#797111[megatron, vllm, env] feat: add native NVFP4 training and rollout with live refit
huggingface/peft#383211FIX strict_adapter_check for load_adapter and from_pretrained
vllm-project/vllm#6160710[XPU] Add Qwen4Exp (Qwen3.8-Flash-Next) model architecture support
unslothai/unsloth#1599310Studio: decode with CUDA graphs by default in the transformers chat backend
unslothai/unsloth#1634410Studio: wait on worker exit instead of fixed sleeps when switching models
unslothai/unsloth#1403910Unsloth Studio (AMD): leave an iGPU out of the Train page VRAM total and the default GPU pick (blocks dual R9700 testing)
deepspeedai/DeepSpeed#860210feat(rollout): segment-based kernel injection for hybrid-engine greedy decode
sgl-project/sglang#3735410[Feat] [Apple] mlx: Add an architecture-independent ATen-to-MLX lowering layer
vllm-project/vllm#371919[Bug]: Benchmark tuning MoE kernel config fails due to triton 3.6
vllm-project/vllm#643409[Bugfix][MoE] Preserve invalid expert IDs in DeepEP LL round-robin mapping
vllm-project/vllm#537399[Bugfix][Parser] Terminate streamed tool arguments the flush cannot extend
unslothai/unsloth#160419Gemma 3: generate on the non-flex attention path instead of flex_attention's fallback
unslothai/unsloth#160739Flex attention on torch < 2.14: fitting tiles for short decode calls so a batch-size change does not leave the process on eager attention
unslothai/unsloth#161979Reset the inference stance counter in for_inference
sgl-project/sglang#469539[Feat] weight_cache: Add the vmm_fd transport
sgl-project/sglang#457929[Feat] [NPU] multimodal: Add JPEGD + VPC image decoding
huggingface/peft#35079PeftMixedModel.disable_adapter does not preserve nested or pre-disabled state
vllm-project/vllm#561168[Quantization] Support ModelOpt layer-wise mixed KV cache
vllm-project/vllm#381228[Bug]: Qwen 3.5 fails to load from GGUF
vllm-project/vllm#563278[LoRA][Model] Add LoRA support for GLM-5.3-Flash
sgl-project/sglang#396828[Feat] diffusion/model: support HiDream-O1-Image
verl-project/verl#71358[megatron] feat: support packed cu_seqlens
huggingface/peft#40698fix: reuse the 4-bit quantization scale when unmerging LoRA
vllm-project/vllm#634067[Core][CUDA Graph] Separate prefill (token) and decode (request) CUDA graph capture sizes
vllm-project/vllm#641977[Feature] Support per-model Hugging Face and ModelScope sources
vllm-project/vllm#626677[Frontend] Support sanitize_control_tokens in chat templates
vllm-project/vllm#615427[Bugfix][Tool Parser] Preserve whitespace in HYV4 argument names
vllm-project/vllm#562187[Bugfix][Model] Don't carry realtime context into single-turn prompt models
sgl-project/sglang#279837[Fix] hf_transformers: Apply rope overrides to text_config for VLM-format models
huggingface/lerobot#46267[Bug/Feature Req/Fix?] DAgger rollout strategy cannot record full task-attempt episodes with HIL corrections
vllm-project/vllm#475486[Bug]: Combination of sleep mode and speculative decoding cause crash (ROCm / MI250)
vllm-project/vllm#526306[Performance]: ~18% pooling throughput regression starting with torch 2.11 (vLLM v0.20.0+)
vllm-project/vllm#631706[Bug]: GLM-5.x (DSA) under pipeline parallelism: reuse layers at a stage start read stale top-k indices
vllm-project/vllm#642826[Model] Add Nemotron-3.5-Transcribe
vllm-project/vllm#372206[Bugfix] Consolidate Gemma2/3 GGUF fixes for correctness on Blackwell
vllm-project/vllm#434086[Feature] Tolerate malformed Harmony streams.
unslothai/unsloth#163176[Bug] Unsloth North Mini Code giving Stream interrupted error when using Bearer Token and OpenAIAPI with OpenCode
unslothai/unsloth#159346cuDNN SDPA: keep known-bad decode and NaN-dQ shapes off cuDNN before torch 2.14 / 2.15
unslothai/unsloth#160496Flex attention: small shared-memory tiles for short no-grad calls so one short prompt does not leave the process on eager attention
sgl-project/sglang#477856[Perf] [NVIDIA] qwen3: Fuse prefill QK norm and RoPE with contiguous Q
sgl-project/sglang#330596[Fix] [AMD] dsa: Align Aiter paged rows with DP-attention metadata
sgl-project/sglang#472716[Feat] Mistral-Large-4: add model support
huggingface/lerobot#46346perf(smolvla): static immutable prefix KV cache and torch.compile support (4.2x speedup)
verl-project/verl#83986[rollout, vllm] fix: keep vLLM's MXFP4 MoE selection on Hopper instead of forcing deep_gemm
vllm-project/vllm#612675[Bugfix][LoRA] respect `attention_k_eq_v=True` QKK LoRA layers
vllm-project/vllm#506165[Bug]: runai_streamer multi-node: aux (non-tensor) files are not pulled on peer nodes
vllm-project/vllm#650345[Bug]: Quark `exclude` entries written as weight names (e.g. `mtp.fc.weight`) are silently never matched, breaking MTP speculative decoding
vllm-project/vllm#459785[Bugfix] Honor override_generation_config for special tokens
vllm-project/vllm#386065[Bug]: KV block corruption under rapid LoRA adapter alternation
vllm-project/vllm#371595[Bug]: vLLM crashed with V100 by running zai-org/GLM-OCR
vllm-project/vllm#372615[Bug]: KeyError when use lmcache layerwise mode in vllm 0.14.0
vllm-project/vllm#615245fix(renderers): preserve special tokens when truncation_side matches tokenizer default
vllm-project/vllm#621625[Feature][KV Offload] Enable canonical offload for chunked-local attention
unslothai/unsloth#159765Decode CUDA graphs: opt-in whole-step capture for batch 1 (UNSLOTH_DECODE_CUDA_GRAPHS_FULL=1)
unslothai/unsloth#162655Studio RAG: count chunking tokens in batches
unslothai/unsloth#162665Studio export: upload a reused folder's fresh files instead of saving the model twice
deepspeedai/DeepSpeed#87775[RFC] Backend-pluggable low-precision training for Dense Linear and AutoEP grouped GEMM (Ascend HiFloat8)
sgl-project/sglang#372165[Bug] remote_instance weight loading fails with --pp-size 2 โ€” TransferEngine info keyed by tp_rank
huggingface/diffusers#151315Flux2KleinInpaintPipeline: `padding_mask_crop` pastes a center crop of the result instead of resizing it
huggingface/lerobot#51635fix(policies): let LingBot-VLA v2, the LaWAM VLM stage and the GR00T N1.7 backbone pass torch.export
huggingface/peft#38305OFT module_dropout has had no effect since #2575: the MultiplicativeDropoutLayer is created but never called
huggingface/peft#39425FIX Support MoE expert target_parameters with TP/EP
huggingface/trl#76185SDFT and SDPO build the teacher in fp32 with the default attention when the model is passed as an instance
vllm-project/vllm#650054[Docs][Model] Add Parakeet TDT transcription support via plugin
vllm-project/vllm#639094[Bugfix][Spec Decode] Propagate dictionary hf-overrides to speculative draft models
vllm-project/vllm#505774[Bug]: AxionML/Gemma-4-12B-NVFP4 fails with shape mismatch in flashinfer_scaled_fp4_mm on RTX 5060 Ti (sm_120)
vllm-project/vllm#648744[Bug]: [ROCm][GLM-5.3-Flash]: "Workspace is locked" AssertionError in rocm_fp8_paged_mqa_logits under high decode concurrency (glm5next AMD indexer reserves decode logits via torch.empty, not the workspace)
vllm-project/vllm#649594[Model] Add LiquidAI LFM3 (Lfm3ForCausalLM)
vllm-project/vllm#643264[Bug]: DelegatingParser.parse ignores model_output_token_ids, so markers written as plain text become tool calls or end reasoning (non-streaming)
vllm-project/vllm#637324[Bug]: Gemma 4 + online --quantization fp8: image requests crash since v0.23.0 (FP8-cast pixel values reach quantized patch_embedder.input_proj); v0.22.0 works
vllm-project/vllm#371854[Bug]: The 2-bit model repeatedly outputs !!!!, while Transformers works correctly.
unslothai/unsloth#163654feat(studio): system-wide voice typing with on-device grammar
unslothai/unsloth#162934SFT: answer the pre-tokenized cap check from Arrow lengths when they prove it
unslothai/unsloth#161894Import: scan installed distributions once while transformers and TRL import
unslothai/unsloth#162034RL: read package versions once per trainer patch pass
unslothai/unsloth#162434Studio: chat with LFM3-Audio-2B (text and speech in, text out)
hiyouga/LlamaFactory#103884fix: add None check for feature_extractor in Gemma4Plugin audio processing
deepspeedai/DeepSpeed#85014Fix ZeRO-3 checkpoint tag selection during weight reconstruction
sgl-project/sglang#477544[Fix] multimodal: Make CUDA VMM feature transport opt-in
sgl-project/sglang#411104[Feature] Qwen3.5forsequenceclassification model is supported
sgl-project/sglang#357724[Bug] Qwen3-VL vision features diverge from Transformers/vLLM in v0.5.17 on fine-grained grounding
sgl-project/sglang#471034[Perf] scheduler: Add opt-in stop-string tail reuse and prefix pruning
sgl-project/sglang#444354[Feat] [NPU] multimodal: Add opt-in cross-request ViT image batching
sgl-project/sglang#347404[Bug] SGLANG_SIMULATE_ACC_LEN silently degrades detokenization to O(nยฒ) โ€” `predict.fill_(100)` emits a byte-fallback token, and the `endswith("\ufffd")` commit gate then never advances the incremental-detokenization offsets
huggingface/lerobot#51484[Performance] pi0.5: the denoise loop is 86% of a bf16 chunk; static prefix KV + CUDA-graph denoise?
huggingface/lerobot#51064fix(pi0): mask padded actions in the loss
ggml-org/llama.cpp#314713mtmd : add Nemotron 3.5 Transcribe support
ggml-org/llama.cpp#313433gguf-py : fix crash when extra_special_tokens is not a dict
vllm-project/vllm#480763[Performance]: Triton attention drops long-context decode from 3D split-KV to 2D at batch โ‰ฅ ~12, ~doubling ITL
vllm-project/vllm#490313[Bug]: FlashInfer B12x W4A16 runtime repacking retains original and packed MoE weights, causing startup OOM
vllm-project/vllm#505573[Bug]: MiniMax-M3-NVFP4 illegal memory access
vllm-project/vllm#504333[Performance][MooncakeStoreConnector] Overlapping async cold-prefix loads cause duplicate Gets into distinct GPU blocks
vllm-project/vllm#466113[Bug]: No available shared memory broadcast block found in 60 seconds. This typically happens when some processes are hanging or doing some time-consuming work (e.g. compilation, weight/kv cache quantization).
vllm-project/vllm#641193[Fast Start] Support ModelOpt CUTLASS block FP8 cache restarts
vllm-project/vllm#592063[Transformers backend] Migrate ModernBERT embedding, reranker and token classification models
vllm-project/vllm#613203Log EAGLE draft rope base at startup; warn on silent default
vllm-project/vllm#639563[Perf][DSA] Token-sharded sparse-MLA prefill under TP; remove indexer row sharding
vllm-project/vllm#583793[Bugfix][Multimodal] Reject empty Qwen2-VL video samples
vllm-project/vllm#648823[Bug]: Multimodal video loaders ignore --media-io-kwargs num_frames
vllm-project/vllm#645363[Feature][Structured Output] Support continued assistant prefixes
vllm-project/vllm#621603[Bugfix][KV Offload] Deduplicate replicated GQA heads in canonical storage
unslothai/unsloth#162463Tests: gpt-oss dense-expert reload fixture drops replaced experts' quant state
unslothai/unsloth#161783[Bug] Studio: image generation returns 404 after a managed account loads a cached gated model
unslothai/unsloth#163103Studio diffusion: change LoRA weights without reloading the adapter files
unslothai/unsloth#161863[Bug] [Critical Regression / Severity: Incident-level bug] Isolated install via UNSLOTH_STUDIO_HOME silently overwrites existing baseline venv
unslothai/unsloth#163733Notebooks CI: refresh the Colab oracle snapshots to the current GPU runtime
hiyouga/LlamaFactory#108753[export] keep exported tokenizer_config loadable with transformers<5
deepspeedai/DeepSpeed#89803[Hybrid Engine] Require an explicit accelerator opt-in for graph capture
deepspeedai/DeepSpeed#89393feat(rollout): non-greedy sampling on hybrid engine graph decode path
deepspeedai/DeepSpeed#89333[Hybrid Engine] Enable graph capture for HybridEngine on non-CUDA accelerators
deepspeedai/DeepSpeed#86593Add deepspeed.models.glm_moe_dsa: TileLang sparse attention for GLM-5.2
deepspeedai/DeepSpeed#87393Overlap AutoEP's expert weight gradients with the DeepEP backward combine
deepspeedai/DeepSpeed#89593Honor explicit dtype in ZeRO tensor constructors
sgl-project/sglang#473673[Feat] [NPU] attention: support mixed-chunk on the ascend attention backend
sgl-project/sglang#471293[Perf] [NVIDIA] muse-spark: Add B200 runtime optimizations
sgl-project/sglang#477693[Feat] multimodal: Add bounded GPU feature publication admission
sgl-project/sglang#377533[Do not merge - need to clean] GLM-5.3-Flash (glm5_next) LoRA serving + glm53next port rebased
sgl-project/sglang#414033[Benchmark] Add synthetic audio input to bench_serving
sgl-project/sglang#464833[Perf] [NPU] attention: Avoid CPU sequence length synchronization for Kimi K3
sgl-project/sglang#448133[Fix] deepseek_v4: route unquantized BF16/F16/F32 checkpoints off the MXFP4 expert path
sgl-project/sglang#371833AMD MI308X SGLang GLM-5.3-Flash ValueError: The checkpoint you are trying to load has model type `glm5_next` but Transformers does not recognize this architecture.
huggingface/diffusers#151243Quantizer configs are mutated in place, leaking modules_to_not_convert across loads
huggingface/diffusers#151443Fix/flux lora quantization compute embeddings
huggingface/diffusers#151483[core] Enforce minimum supported versions in availability checks
huggingface/diffusers#151473Fix replicated parameter dtypes during tensor-parallel loading
huggingface/lerobot#51553fix(policies): load FAST action tokenizers saved by transformers 4
huggingface/lerobot#47783feat(policies): load GR00T fine-tunes without NVIDIA's base model
huggingface/lerobot#51523perf(pi05): share one read-only prefix cache across the denoise steps
huggingface/lerobot#51513feat(fastwam): add Streaming-WAM to Fast-WAM (2.9x faster warm inference on L40S, no pause at chunk refresh)
huggingface/peft#40503fix: report global dims of tensor-parallel LoRA adapters
huggingface/peft#40673Error message for unmatched target_modules does not reveal that patterns are matched with re.fullmatch
huggingface/peft#38573PeftModelForQuestionAnswering silently drops `token_type_ids` and `position_ids`
ggml-org/llama.cpp#314562gptneox: FFN uses tanh GELU (LLM_FFN_GELU) but GPT-NeoX/Pythia are trained with exact erf GELU
vllm-project/vllm#633692[Multimodal] Add TorchCodec image decoding backend
vllm-project/vllm#590582[Bugfix][MoRIIO] Preserve WRITE receive mappings after consumer abort
vllm-project/vllm#644122[Bugfix][Frontend] Pool HF fast tokenizers in all renderers
vllm-project/vllm#645432[Bugfix][Model] Fix dots_ocr vision tower QKV split for TP > 1
vllm-project/vllm#623952[Feat][2/2] Add Transformers TPSP dense baseline support
vllm-project/vllm#475282[Bug]: DeepSeek-V4-Pro (DeepseekV4ForCausalLM, scale_fmt=ue8m0 / FP4 MoE) produces garbled / degenerate output under tensor parallelism (TP), while data parallelism + expert parallelism (DP+EP) works correctly
vllm-project/vllm#627522[Performance]: Avoidable preemption at hybrid Attention/Mamba decode block transitions
vllm-project/vllm#618882[Distributed] Opt-in FP16 partial sums for tensor-parallel row-parallel layers
vllm-project/vllm#636362[Bugfix][Transformers] Infer mm encoder dynamic_arg_dims from forward signature
vllm-project/vllm#650212[Doc][ROCm] Add verified MI300X and MI325X entries to the Qwen3.8-27Bโ€ฆ
vllm-project/vllm#606082[Bug]: Exllama GPTQ and RDNA3 W4A16 kernels silently produce wrong output when `group_size % 32 != 0` (reproduced on CUDA sm_120 and ROCm gfx1100)
vllm-project/vllm#561802Failed: Cuda error /workspace/csrc/custom_all_reduce.cuh:164 'invalid argument'
vllm-project/vllm#642392[Bug]: Timed-out overflow dequeue returns the wrong message
vllm-project/vllm#649682[WIP][Feature] Qwen GDN BI + MTP + prefix cache integration
vllm-project/vllm#613532[Bugfix][Spec Decode] Fix DFlash capture on FLEX_ATTENTION and honor drafter enforce_eager
vllm-project/vllm#473262[Bug]: Incoherent output depending on tensor-parallel-size during inference on Leanstral (Mistral Small 4 family)
vllm-project/vllm#648992[Bugfix][Multimodal] Honour explicit num_frames in Qwen-VL video loaders
unslothai/unsloth#163402CLI: inference and chat cannot take an image, so a VLM can be trained but not used
unslothai/unsloth#163472CLI: inference and chat take images for vision models
unslothai/unsloth#162122Studio: check the weight index of a module path that only reads as the root
unslothai/unsloth#162422Studio: LFM3-Omni audio input, export, LoRA defaults and Stop for custom-code generate
deepspeedai/DeepSpeed#85022Restore full training state in checkpoint writer engines
deepspeedai/DeepSpeed#87452AutoEP: opt-in skip of the recomputed DeepEP combine
sgl-project/sglang#478302[Bug] Hybrid Mamba model + EAGLE: KV-cache sizing error says to raise --mem-fraction-static when the Mamba state cache is what fills memory
sgl-project/sglang#477702[Perf] DFlash: Use CUDA graphs for draft and idle verify under DP attention
sgl-project/sglang#477292[Feat] qwen3.5: Support native sequence classification
sgl-project/sglang#425512[Feature] Support API keys from secret environment variables
sgl-project/sglang#425502[Fix] Reject removal of an external corpus while its load is pending
sgl-project/sglang#425492[Fix] Preserve Gemma4 null and exponent tool argument types
sgl-project/sglang#373492[Guide] Setup and Troubleshooting for SGLang on V100
sgl-project/sglang#431022[Feat] [NPU] qwen4: Support Qwen3.8-Flash-Next W8A8 inference with QSA on Ascend A2
sgl-project/sglang#403562[Fix] [Ascend] bench_serving: Correct Kimi K2.5 image token counts
sgl-project/sglang#474162[Fix] metrics: Correct MFU/MBU bandwidth overcounting and GraniteMoE AutoTokenizer fallback
sgl-project/sglang#370522[Bug] Qwen3.8-Flash-Next + NEXTN full decode graph: repeated dual-rank invalid-probability asserts on GB10
sgl-project/sglang#370972[Bug] Pretokenized image requests can crash GLM MRoPE with a stale retokenized mask
huggingface/diffusers#151612DreamLite: disabling attention slicing replaces the GQA processor and breaks inference
huggingface/lerobot#50992[Discussion] Has the bf16 update-erasure paper been evaluated or reproduced in LeRobot?
verl-project/verl#84142[model] perf: build Qwen3-VL/Qwen3.5 pos-embed interpolation tensors without Python lists
huggingface/peft#40862Fix broken examples on transformers 5.x (target_modules, dataset IDs, warmup_ratio)
huggingface/peft#39382FIX Reshard on disable_adapter() exit only when needed
huggingface/trl#80062Fail CI test jobs on self-hosted runners when pytest does not complete
huggingface/trl#80582[WIP] Quantize the tiny DeepSeek-V3 and GPT-OSS models like their references (FP8, MXFP4)
huggingface/trl#80502`parse_response` puts Qwen3.5 tool calls into `reasoning_content` when the model skips the prefilled `<think>`
huggingface/trl#79652[BUG] GRPO/RLOO continuous batching casts the trained weights to bf16 in place
vllm-project/vllm#650751[Bug]: Bi-encoder scoring token limits fail with MistralTokenizer
vllm-project/vllm#650741[Bug]: MCP tool discovery drops tools after the first tools/list page
vllm-project/vllm#650811[CI Failure]: pre-commit mypy on main - process_weights_after_loading redefined in vllm/models/muse_spark/amd/model.py
vllm-project/vllm#650711[Bug]: CustomDataset converts numeric-string prompts to integers before tokenization
vllm-project/vllm#650721[Bug]: ASRDataset uses channel count as duration when filtering stereo audio bytes
vllm-project/vllm#639061[Bug]: FunctionGemma converts escape-delimited string literals into numbers, booleans and null
vllm-project/vllm#650481[CI Failure]: MI300 MoE Kernels Multi-GPU - MoriEPv2 combine block_num 304 exceeds xdb_flag_slots 256 in test_modular_kernel_combinations.py
vllm-project/vllm#507921[Bug]: OverflowError crash in `prompt_logprobs` decode path on PD-disaggregated (external KV connector) setup
vllm-project/vllm#650351[Bug]: Anthropic Messages stream reports the whole prompt as input_tokens in message_start
vllm-project/vllm#634871[Bugfix] Propagate EngineDeadError on AsyncLLM output_handler cancellation
vllm-project/vllm#650151[Bugfix][Tool Parser] Keep granite4 streamed text between tool calls
vllm-project/vllm#650021[Model] Add NVIDIA Nemotron-3-Diarization (offline speaker diarization)
vllm-project/vllm#648091[Bug]: `raw_logprobs` are computed after the structured-output grammar mask
vllm-project/vllm#649561[Bug]: kv_cache_dtype="turboquant_3bit_nc" crashes on Phi-2 (head_size=80) โ€” supports_head_size() claims "any positive value", but the Hadamard rotation is built at the next power of two (128ร—128) and matmul'd with an 80-dim key
vllm-project/vllm#649611[Bugfix][Frontend] Convert Responses content parts to chat parts
vllm-project/vllm#649531[Bug]: `seed=-1` passes config validation and kills engine init inside `np.random.seed` ("Seed must be between 0 and 2**32 - 1") โ€” while `SamplingParams` already treats `-1` as "no seed"
vllm-project/vllm#648851[Bugfix][Pooling] Enable Matryoshka dimensions for EmbeddingGemma 2
vllm-project/vllm#649031[Bugfix][MoE] Avoid double router weighting in TRTLLM FP8 experts
vllm-project/vllm#648451[Bug]: EmbeddingGemma 2 rejects `dimensions` by default although the model supports Matryoshka (128/256/512/768)
unslothai/unsloth#164071UNSLOTH_PARALLEL: one flag and one root mesh for dp / tp / pp / cp, pipeline parallel from the environment, and a capability registry
unslothai/unsloth#164161Packed mRoPE attention: no graph break on warm packed rows
unslothai/unsloth#161691[Bug] Sandbox error that ejects the model?
unslothai/unsloth#164061Pipeline parallel: carry the Qwen3-MoE router auxiliary loss across stages
unslothai/unsloth#164101Block compile: install the router_logits capture hooks before the first block compiles
unslothai/unsloth#163311CLI: no way to pass trust_remote_code, so models with custom code cannot load
unslothai/unsloth#163321CLI: let train load models that need trust_remote_code
unslothai/unsloth#162211[Bug] ShareGPT image lists fail after passing dataset format checks
unslothai/unsloth#162941[Bug] Studio video: MiniMax-H3 on Auto attention crashes with "No available kernel" on Windows (torch without flash attention)
unslothai/unsloth#156321Block Swap blocker: Qwen3.8-Flash-Next-FP8 can't train with offload_layers, its FP8 n-gram table loads without its weight_scale
unslothai/unsloth#162521Tests: compare the disabled 4-bit fused LoRA forward with dequantize + linear, not bitsandbytes' fused GEMM
unslothai/unsloth#161041Saving: write the source tokenizer_class instead of TokenizersBackend
unslothai/unsloth#161021Saving: drop the training truncation and padding from saved tokenizers
hiyouga/LlamaFactory#108681enable fp16 in dtype inference on Intel XPU
deepspeedai/DeepSpeed#89501Fix torch-latest workflow unit failures
deepspeedai/DeepSpeed#88511[AutoTP] Support frozen AutoTP models and manage AutoTP mode per model
deepspeedai/DeepSpeed#88431Keep the parameters a model names for fp32 in fp32 under ZeRO-3
sgl-project/sglang#463031[Perf] diffusion: fix 6 quantization-layer overheads (NVFP4 scale conversion, W8A16, BnB, Nunchaku)
sgl-project/sglang#462481[Feat] hicache: Add unified Engram store and support Qwen4-Exp
sgl-project/sglang#478411[Fix] transcription: Serve /v1/audio/transcriptions for Qwen3-Omni and Gemma 4
sgl-project/sglang#457981[Fix] llava: Load LLaVA-1.5 CLIP tower
sgl-project/sglang#399721[Model] Add BerryLM
sgl-project/sglang#376481[Bug][ROCm] GLM-5.3-MXFP4 TP8 + EAGLE GPU memory fault under high-concurrency long-context load
sgl-project/sglang#376461[Bug] --enable-flashinfer-pure-allreduce is not available on 0.5.18
sgl-project/sglang#376451[Bug] glm45 reasoning parser returns the whole answer as reasoning_content (content empty) when the model skips thinking and stops on EOS
sgl-project/sglang#477861[Feat] multimodal: Share CPU preprocessing artifacts across tokenizer workers
sgl-project/sglang#470601[Fix] observability: Snapshot STANDALONE metrics before settlement
sgl-project/sglang#477461[Fix] diffusion: SenseNova-U1: keep think tokens in attention, fix 4.x import and vision_model init
sgl-project/sglang#476491[Fix] scoring: Return token_logprobs for setwise CausalLM scoring
sgl-project/sglang#475921[CI] ci: Preload `sglang.test.test_utils` in CPU fork workers
sgl-project/sglang#372831[Bug] A non-object entry in a tool-call array discards the valid calls beside it
sgl-project/sglang#372151[Bug] --dp 8 intermittently fails with TCPStore EADDRINUSE on single-node 8ร—H800
sgl-project/sglang#371861[Bug] Qwen3.5 visual encoder weights dropped by PP layer-range filter in load_weights
huggingface/diffusers#151211Fix several quantizer config bugs (AutoRound override, ModelOpt JSON crash, single-file pre_quantized, config mutation, Nunchaku group_size, dataclass eq/repr)
huggingface/diffusers#151631Flux and LTX pipelines do not forward the generator to stochastic scheduler steps
huggingface/lerobot#47751feat(policies): load pi0, pi0.5 and pi0-FAST checkpoints without first building random weights
huggingface/lerobot#51461fix(eo1): give the slow Qwen2.5-VL image processor CPU images
huggingface/lerobot#51621fix(policies): let EO-1, WALL-X and Wall-OSS 0.5 pass torch.export
huggingface/lerobot#51341feat(policies): add Cosmos3
verl-project/verl#84041[hardware, megatron, doc] feat: upgrade Ascend 950 to Megatron-Core 0.18 and add MXFP8 GRPO example
huggingface/peft#40961Update FSDP docs
huggingface/peft#40941reduce_intruder_dimension gives wrong LoRA factors for fan_in_fan_out (Conv1D) and Embedding layers
huggingface/peft#40901layers_to_transform without layers_pattern matches nothing when the layer list is at the model root
huggingface/peft#40881[v1 #4071] Disjoint auxiliary trainable tokens activate only the first adapter
huggingface/peft#40931CI: Bump the third-party-actions group with 5 updates
huggingface/trl#80401Add experimental DecisionTrainer
huggingface/trl#80541Allow `<tool_call>` to close reasoning region in `qwen3_5_template`
huggingface/trl#80521SFTTrainer silently skips ModernBERT Decoder's prediction head
ggml-org/llama.cpp#309220mtmd : resize Gemma 4 images without letterbox padding
vllm-project/vllm#650730[Bug]: Duplicate MCP server names advertise tools from a different execution server
vllm-project/vllm#633000[Rust Frontend] Complete partial cached HF snapshots when online
vllm-project/vllm#650550[Performance]: MoE top-k gating PDL launch (#59737) slows CUDA graph replay on pre-Hopper GPUs
vllm-project/vllm#639830[Bug]: FunctionGemma tool parser drops every non-string argument (numbers, booleans, arrays, objects)
vllm-project/vllm#645410[Bug]: dots.ocr produces identical "blurry image" output for every region with tensor-parallel size 2
vllm-project/vllm#644180[RFC]: Load plan โ€” an engine-neutral API for sharded weight transfer
vllm-project/vllm#649540[Bug]: `--attention-backend TORCH_SDPA` passes config validation, then dies with "Backend TORCH_SDPA must be registered before use" โ€” a ViT-only tag offered as a valid option, whose `ValueError` bypasses the "Selected backend ... is not valid" path
vllm-project/vllm#624700[Bugfix][LoRA] Scale embedding LoRA deltas by the base layer's embed_scale
unslothai/unsloth#164130Studio: keep a generated chat title in the conversation's language
unslothai/unsloth#163770Studio: apply the mapped chat template once per loaded tokenizer
unslothai/unsloth#157650Studio: show measured transcription progress
unslothai/unsloth#162790Ascend: opt-in eager bootstrap for Torch 2.6/Triton 3.5 on ARM64
unslothai/unsloth#156330Block Swap blocker: offload_layers never engages on Qwen3.8-Flash-Next, the CPU-kept n-gram table makes the map look like two devices
unslothai/unsloth#151590Block Swap: get_peft_model crashes on MoE models with offload_layers (evicted expert weights read as empty)
hiyouga/LlamaFactory#109290[data] Preserve literal placeholders in user content during template encoding
hiyouga/LlamaFactory#108760[xpu] support profiler, module profiler and env report on Intel GPU
hiyouga/LlamaFactory#109270[chat] pass mm_token_type_ids to generate for mrope models
hiyouga/LlamaFactory#109260[data] avoid mutating Qwen video frame inputs
hiyouga/LlamaFactory#109240feat(data): optionally drop overlong SFT examples
hiyouga/LlamaFactory#109230feat(webui): support private Hub export repositories
hiyouga/LlamaFactory#108490[data] guard Qwen2-VL media tokens against SFT truncation
hiyouga/LlamaFactory#109180[npu] Skip unnecessary gradients in NPU fused MoE backward
hiyouga/LlamaFactory#109220fix: handle a single validation sample after CLI parsing
deepspeedai/DeepSpeed#89320[REQUEST][Hybrid Engine] Enable graph capture for HybridEngine on non-CUDA accelerators
deepspeedai/DeepSpeed#89620Fix AutoTP multi-query GPTBigCode training at TP > 1
deepspeedai/DeepSpeed#89610[BUG] AutoTP multi-query GPTBigCode training fails at TP > 1
sgl-project/sglang#438640[Bug] gemma-4-31B-it: audio request crashes the scheduler; /get_model_info misreports has_audio_understanding: true
sgl-project/sglang#477310[Perf] imports: Keep light modules off HF patches, model configs, and the scheduler
sgl-project/sglang#477100[Bug] SM120: fused-A GEMM eligibility accepts K % 256 but the auto-selected CuTe DSL kernel asserts K % 1024 (hidden_size=3584 crashes)
sgl-project/sglang#478090[Fix] attention: Honor custom masks past the causal tile in Triton extend
sgl-project/sglang#477990[Fix] models: Phi-3-4k ignores its 2047-token sliding window (output breaks past 2k tokens)
sgl-project/sglang#477590[Bug] Video input data still leak into logs at --log-requests-level 0/1
sgl-project/sglang#412740[sglang-miles] Recover disk-delta checkpoints safely across resets and retries
sgl-project/sglang#477010[Feat] activevln: Add differentiable actor and SRT weight export
sgl-project/sglang#474440[Bug] NemotronHConfig rejects transformers 5 layer types `linear_attention` / `full_attention` in config.json
sgl-project/sglang#472230[Fix] constrained: Share one compile among concurrent requests for a grammar key
huggingface/diffusers#151680Bump the actions group across 1 directory with 16 updates
huggingface/diffusers#151670UniPC and DPM-Solver with use_flow_sigmas condition the model on integer timesteps while integrating with float sigmas
huggingface/diffusers#151660[training scripts] `--logit_mean` biases timesteps in the opposite direction to the SD3 paper (positive m oversamples low noise)
huggingface/diffusers#151640Add Kandinsky S2V 1.0 (speech-to-video) as a modular pipeline
huggingface/diffusers#151450flux_lora_quantization example: compute_embeddings.py crashes on current transformers
huggingface/lerobot#51940feat(sim): NVIDIA RoboLab benchmark on Isaac Lab
huggingface/lerobot#51890`lerobot-train` profile on one H200 (ACT, SmolVLA, Diffusion): where step time and memory go
huggingface/lerobot#51880perf(pi0): share one read-only prefix cache across the denoise steps
huggingface/lerobot#50670perf(smolvla): precompute cross-attention expert K/V once per chunk
verl-project/verl#80970[fsdp, model] fix: gather DTensor LM heads for fused kernels
huggingface/peft#40850Fix modules_to_save selection when merging an inactive adapter
huggingface/peft#40890FEAT healthcheck: warn when a randomly initialized task head is not saved
huggingface/trl#80560GRPO/RLOO under tensor parallelism: TP ranks sample different completions, and the prompt sampler is sized for every process
huggingface/trl#80510Docs: smol-course links in the SFT, DPO and ORPO trainer pages return 404
+
actioncomponentwhatthreads
๐Ÿ”ดfix in transformersmodel:QwenQwen3-VL vision features diverge from Transformers in fine-grained groundingsgl-project/sglang#35772
๐Ÿ”ดfix in transformerstokenizersGGUF conversion crashes when extra_special_tokens is not a dictggml-org/llama.cpp#31343
๐Ÿ”ดfix in transformersmodel:GemmaGemma 4 + FP8 quantization causes image request crashes due to FP8-cast pixel valuesvllm-project/vllm#63732
๐Ÿ”ดfix in transformersmodel:GLMGLM-5.x under pipeline parallelism reuses layers with stale top-k indicesvllm-project/vllm#63170
๐Ÿ”ดfix in transformersmodel:DeepSeekDeepSeek-V4 checkpoints with FP8 config incorrectly route through MXFP4 expert pathsgl-project/sglang#44813
๐Ÿ”ดreply/guide usersfrom_pretrainedfrom_pretrained fails to load FAST action tokenizers from transformers 4huggingface/lerobot#5155
๐Ÿ”ดreply/guide usersmodel:QwenQwen3.8-Flash-Next missing from_pretrained generation configNVIDIA-NeMo/Automodel#4451
๐Ÿ”ดreply/guide usersmodel:GLMGLM-5.3-Flash missing pretrained generation config in exportNVIDIA-NeMo/Automodel#4576
๐ŸŸ fix in transformersquantizationBitsAndBytes 4-bit unmerge loses quantization scale, drifting base weightshuggingface/peft#4069
๐ŸŸ fix in transformersmodel:QwenQwen3-MoE and Gemma 4 forced float32 patches not undone when flag turns offunslothai/unsloth-zoo#2325
๐ŸŸ fix in transformerstokenizerstokenizer_config.json with TokenizersBackend class breaks loading in transformers<5hiyouga/LlamaFactory#10875
๐ŸŸ reply/guide usersgenerateSGLang detokenization degrades to O(nยฒ) with SGLANG_SIMULATE_ACC_LENsgl-project/sglang#34740
๐ŸŸ reply/guide usersmodel:GemmaGemma4Plugin audio processing fails due to missing None check in feature_extractorhiyouga/LlamaFactory#10388
โšชreply/guide usersfrom_pretrainedPeftModelForQuestionAnswering silently drops token_type_ids and position_idshuggingface/peft#3857
โšชreply/guide userstokenizersError message for unmatched target_modules doesn't reveal re.fullmatch usagehuggingface/peft#4067
+
All 466 matching threads +
threadcommentstitle
embeddings-benchmark/mteb#4842217[Research Project] MOEB: Massive Omni Embedding Benchmark
huggingface/peft#2310174Comparison of Different Fine-Tuning Techniques for Conversational AI
vllm-project/vllm#50576125[Feature]: SM8x (Ampere A100/A800) support for DeepSeek-V4-Flash / DeepSeek-V4-Flash-0731 DSpark
unslothai/unsloth#1473040Studio: Flows, a node editor for AI pipelines, story games and 2D worlds
vllm-project/vllm#5329830[Bugfix][KV Connector] Restore state after KV load rejection
huggingface/lerobot#500830feat(policies): add native Long-WAM training and inference
sgl-project/sglang#4609127[Feat] usage: align prompt and cache accounting across APIs
sgl-project/sglang#4191625[Fix] sampling: Make top-k/top-p filtering consistent across sampling paths
huggingface/optimum#35123Add all available ONNX models to ORTConfigManager
vllm-project/vllm#6370021[Bugfix][MoE] Avoid double router weighting in TRTLLM BF16 experts
vllm-project/vllm#5615520[Core][Attention] Resolve per-layer KV dtype in attention runtime
sgl-project/sglang#4722919[Fix] kernels: Support OOT fused ops under torch.compile
vllm-project/llm-compressor#280819AWQ GQA support
unslothai/unsloth#1607118LoRA: add the rank-r term inside the base GEMM so small updates are not rounded away
verl-project/verl#748317[vllm] feat: Support LoRA for DeepSeek V4 & DeepSeek V4.1 & GLM5.3 Flash
vllm-project/vllm#5594716[Feature][Spec Decode] Add Uno shared-model drafting with a draft-only LoRA adapter (Model Runner V2)
vllm-project/vllm#5731113[Perf][Startup] Move atomic_writer out of weight_utils so the model-info cache save stays light
vllm-project/vllm#3185613[Bug]: Text degeneration / repetition loops with MiniMax-M2.1-NVFP4 on v0.14.0rc1.dev308
sgl-project/sglang#4685813[Feat] lfm3: Add text, vision, audio and Omni support
sgl-project/sglang#3605913[Perf] dsv4: Add native H16 CuTe sparse prefill backend
sgl-project/sglang#3077113[Fix] tokenizer: Isolate and offload regular fast tokenization
unslothai/unsloth#1603212Opt-in fused batch-1 NF4 decode for Llama (UNSLOTH_DECODE_FUSED_NF4=1): 1.6-2.3x per token in eager generate
unslothai/unsloth#1622712Studio MLX: decode reply snapshots incrementally instead of re-decoding every id
NVIDIA/TensorRT-LLM#2125112[https://nvbugs/7011503][fix] Pass max_seq_len to MiniCPM-V 4.6 vision attention metadata
vllm-project/vllm#5969311[ROCm][Perf] Kimi-K3: token-sharded residual stream for long prefills
unslothai/unsloth#711511Add EXL3 (ExLlamaV3) quantization backend: 2/3/4/6/8-bit and MoE support
sgl-project/sglang#4562111[Feat] decisions: Serve StartLux-Decision checkpoints on /v1/systemone
sgl-project/sglang#4750311[Perf] rmsnorm: One SM100-tuned JIT template for plain, fused residual-add and fp8 per-tensor quant
sgl-project/sglang#3322611[Feat] cp: Support FlashInfer prefill context parallelism
sgl-project/sglang#4488311[Feat] [NPU] multimodal: Enable device-side multimodal feature transport via torch_npu IPC
verl-project/verl#797111[megatron, vllm, env] feat: add native NVFP4 training and rollout with live refit
huggingface/peft#383211FIX strict_adapter_check for load_adapter and from_pretrained
NVIDIA-NeMo/Automodel#431411feat(glm5_next): packed CP/PP training and GLM-5.3-Flash kernel backends
vllm-project/vllm#6160710[XPU] Add Qwen4Exp (Qwen3.8-Flash-Next) model architecture support
unslothai/unsloth#1599310Studio: decode with CUDA graphs by default in the transformers chat backend
unslothai/unsloth#1634410Studio: wait on worker exit instead of fixed sleeps when switching models
unslothai/unsloth#1403910Unsloth Studio (AMD): leave an iGPU out of the Train page VRAM total and the default GPU pick (blocks dual R9700 testing)
deepspeedai/DeepSpeed#860210feat(rollout): segment-based kernel injection for hybrid-engine greedy decode
sgl-project/sglang#3735410[Feat] [Apple] mlx: Add an architecture-independent ATen-to-MLX lowering layer
intel/auto-round#254010refine export and format code and provide unify api
vllm-project/vllm#371919[Bug]: Benchmark tuning MoE kernel config fails due to triton 3.6
vllm-project/vllm#643409[Bugfix][MoE] Preserve invalid expert IDs in DeepEP LL round-robin mapping
vllm-project/vllm#537399[Bugfix][Parser] Terminate streamed tool arguments the flush cannot extend
unslothai/unsloth#160419Gemma 3: generate on the non-flex attention path instead of flex_attention's fallback
unslothai/unsloth#160739Flex attention on torch < 2.14: fitting tiles for short decode calls so a batch-size change does not leave the process on eager attention
unslothai/unsloth#161979Reset the inference stance counter in for_inference
sgl-project/sglang#469539[Feat] weight_cache: Add the vmm_fd transport
sgl-project/sglang#457929[Feat] [NPU] multimodal: Add JPEGD + VPC image decoding
huggingface/peft#35079PeftMixedModel.disable_adapter does not preserve nested or pre-disabled state
NVIDIA/TensorRT-LLM#202859[https://nvbugs/6787519][fix] Initialize Nemotron-H-Puzzle configuration correctly
vllm-project/vllm#561168[Quantization] Support ModelOpt layer-wise mixed KV cache
vllm-project/vllm#381228[Bug]: Qwen 3.5 fails to load from GGUF
vllm-project/vllm#563278[LoRA][Model] Add LoRA support for GLM-5.3-Flash
sgl-project/sglang#396828[Feat] diffusion/model: support HiDream-O1-Image
verl-project/verl#71358[megatron] feat: support packed cu_seqlens
huggingface/peft#40698fix: reuse the 4-bit quantization scale when unmerging LoRA
NVIDIA-NeMo/Automodel#42088refactor(speculative): share draft recipe lifecycle in a base class
modelscope/ms-swift#83268sftๅพฎ่ฐƒqwen3.5 ๆž„้€ ่‡ชๅทฑ็š„VQAๆ•ฐๆฎ้›†jsonl ๏ผŒๆŠฅ้”™An error occurred while generating the dataset
embeddings-benchmark/mteb#54008Add public NGNN general text encoder with fixed 512D compression
unslothai/unsloth-zoo#22968Reset the inference stance counter at every generate
vllm-project/vllm#634067[Core][CUDA Graph] Separate prefill (token) and decode (request) CUDA graph capture sizes
vllm-project/vllm#641977[Feature] Support per-model Hugging Face and ModelScope sources
vllm-project/vllm#626677[Frontend] Support sanitize_control_tokens in chat templates
vllm-project/vllm#615427[Bugfix][Tool Parser] Preserve whitespace in HYV4 argument names
vllm-project/vllm#562187[Bugfix][Model] Don't carry realtime context into single-turn prompt models
sgl-project/sglang#279837[Fix] hf_transformers: Apply rope overrides to text_config for VLM-format models
huggingface/lerobot#46267[Bug/Feature Req/Fix?] DAgger rollout strategy cannot record full task-attempt episodes with HIL corrections
ml-explore/mlx-lm#12617Support for Zyphra/ZAYA1-base
unslothai/unsloth-zoo#22437Vendored fla: faster KDA decode, masked KDA output-kernel load, cached backend probe (fla #1415 #1416 #1582 #996)
vllm-project/vllm#475486[Bug]: Combination of sleep mode and speculative decoding cause crash (ROCm / MI250)
vllm-project/vllm#526306[Performance]: ~18% pooling throughput regression starting with torch 2.11 (vLLM v0.20.0+)
vllm-project/vllm#631706[Bug]: GLM-5.x (DSA) under pipeline parallelism: reuse layers at a stage start read stale top-k indices
vllm-project/vllm#642826[Model] Add Nemotron-3.5-Transcribe
vllm-project/vllm#372206[Bugfix] Consolidate Gemma2/3 GGUF fixes for correctness on Blackwell
vllm-project/vllm#434086[Feature] Tolerate malformed Harmony streams.
unslothai/unsloth#163176[Bug] Unsloth North Mini Code giving Stream interrupted error when using Bearer Token and OpenAIAPI with OpenCode
unslothai/unsloth#159346cuDNN SDPA: keep known-bad decode and NaN-dQ shapes off cuDNN before torch 2.14 / 2.15
unslothai/unsloth#160496Flex attention: small shared-memory tiles for short no-grad calls so one short prompt does not leave the process on eager attention
sgl-project/sglang#477856[Perf] [NVIDIA] qwen3: Fuse prefill QK norm and RoPE with contiguous Q
sgl-project/sglang#330596[Fix] [AMD] dsa: Align Aiter paged rows with DP-attention metadata
sgl-project/sglang#472716[Feat] Mistral-Large-4: add model support
huggingface/lerobot#46346perf(smolvla): static immutable prefix KV cache and torch.compile support (4.2x speedup)
verl-project/verl#83986[rollout, vllm] fix: keep vLLM's MXFP4 MoE selection on Hopper instead of forcing deep_gemm
NVIDIA/TensorRT-LLM#208536[None][feat] Classify startup failures, surface their root cause and report them in telemetry
intel/auto-round#24576fix: disable trust_remote_code by default for security
intel/auto-round#25466fix: register copied-through MTP modules in compressed-tensors ignore list
unslothai/unsloth-zoo#22406GLM-5.3-Flash DSA: exact skip, chunked pooled indexer, opt-in flex BlockMask; free failed flex calls' tensors
vllm-project/vllm#612675[Bugfix][LoRA] respect `attention_k_eq_v=True` QKK LoRA layers
vllm-project/vllm#506165[Bug]: runai_streamer multi-node: aux (non-tensor) files are not pulled on peer nodes
vllm-project/vllm#650345[Bug]: Quark `exclude` entries written as weight names (e.g. `mtp.fc.weight`) are silently never matched, breaking MTP speculative decoding
vllm-project/vllm#459785[Bugfix] Honor override_generation_config for special tokens
vllm-project/vllm#386065[Bug]: KV block corruption under rapid LoRA adapter alternation
vllm-project/vllm#371595[Bug]: vLLM crashed with V100 by running zai-org/GLM-OCR
vllm-project/vllm#372615[Bug]: KeyError when use lmcache layerwise mode in vllm 0.14.0
vllm-project/vllm#615245fix(renderers): preserve special tokens when truncation_side matches tokenizer default
vllm-project/vllm#621625[Feature][KV Offload] Enable canonical offload for chunked-local attention
unslothai/unsloth#159765Decode CUDA graphs: opt-in whole-step capture for batch 1 (UNSLOTH_DECODE_CUDA_GRAPHS_FULL=1)
unslothai/unsloth#162655Studio RAG: count chunking tokens in batches
unslothai/unsloth#162665Studio export: upload a reused folder's fresh files instead of saving the model twice
deepspeedai/DeepSpeed#87775[RFC] Backend-pluggable low-precision training for Dense Linear and AutoEP grouped GEMM (Ascend HiFloat8)
sgl-project/sglang#372165[Bug] remote_instance weight loading fails with --pp-size 2 โ€” TransferEngine info keyed by tp_rank
huggingface/diffusers#151315Flux2KleinInpaintPipeline: `padding_mask_crop` pastes a center crop of the result instead of resizing it
huggingface/lerobot#51485[Performance] pi0.5: the denoise loop is 86% of a bf16 chunk; static prefix KV + CUDA-graph denoise?
huggingface/lerobot#51635fix(policies): let LingBot-VLA v2, the LaWAM VLM stage and the GR00T N1.7 backbone pass torch.export
huggingface/peft#38305OFT module_dropout has had no effect since #2575: the MultiplicativeDropoutLayer is created but never called
huggingface/peft#39425FIX Support MoE expert target_parameters with TP/EP
huggingface/trl#76185SDFT and SDPO build the teacher in fp32 with the default attention when the model is passed as an instance
NVIDIA-NeMo/Automodel#45345fix(deepseek-v4): keep the text-only forward CUDA-graph capturable
NVIDIA-NeMo/Automodel#45145perf(moe): add configurable FSDP block prefetch
NVIDIA-NeMo/Automodel#45425feat(glm5_next): train and export the GLM-5.3-Flash MTP layer
NVIDIA-NeMo/Automodel#44545fix(glm5_next): keep layer split and architectures in exported config
modelscope/ms-swift#83145Qwen 3.5 35B A3B SFT example perf issue
modelscope/ms-swift#95935ๅฆ‚ไฝ•้‡ๆž„dataset็š„้ป˜่ฎคgetitemๅ‡ฝๆ•ฐ
vllm-project/llm-compressor#34185Fix distributed oneshot calibrating every rank on the same rows
embeddings-benchmark/mteb#56525fix: video frame timing
unslothai/unsloth-zoo#23255Qwen3-MoE / Gemma 4: undo the forced float32 patches when a later load turns the flag off
vllm-project/vllm#650054[Docs][Model] Add Parakeet TDT transcription support via plugin
vllm-project/vllm#639094[Bugfix][Spec Decode] Propagate dictionary hf-overrides to speculative draft models
vllm-project/vllm#505774[Bug]: AxionML/Gemma-4-12B-NVFP4 fails with shape mismatch in flashinfer_scaled_fp4_mm on RTX 5060 Ti (sm_120)
vllm-project/vllm#648744[Bug]: [ROCm][GLM-5.3-Flash]: "Workspace is locked" AssertionError in rocm_fp8_paged_mqa_logits under high decode concurrency (glm5next AMD indexer reserves decode logits via torch.empty, not the workspace)
vllm-project/vllm#649594[Model] Add LiquidAI LFM3 (Lfm3ForCausalLM)
vllm-project/vllm#643264[Bug]: DelegatingParser.parse ignores model_output_token_ids, so markers written as plain text become tool calls or end reasoning (non-streaming)
vllm-project/vllm#637324[Bug]: Gemma 4 + online --quantization fp8: image requests crash since v0.23.0 (FP8-cast pixel values reach quantized patch_embedder.input_proj); v0.22.0 works
vllm-project/vllm#371854[Bug]: The 2-bit model repeatedly outputs !!!!, while Transformers works correctly.
unslothai/unsloth#163654feat(studio): system-wide voice typing with on-device grammar
unslothai/unsloth#162934SFT: answer the pre-tokenized cap check from Arrow lengths when they prove it
unslothai/unsloth#161894Import: scan installed distributions once while transformers and TRL import
unslothai/unsloth#162034RL: read package versions once per trainer patch pass
unslothai/unsloth#162434Studio: chat with LFM3-Audio-2B (text and speech in, text out)
hiyouga/LlamaFactory#103884fix: add None check for feature_extractor in Gemma4Plugin audio processing
deepspeedai/DeepSpeed#85014Fix ZeRO-3 checkpoint tag selection during weight reconstruction
sgl-project/sglang#477544[Fix] multimodal: Make CUDA VMM feature transport opt-in
sgl-project/sglang#411104[Feature] Qwen3.5forsequenceclassification model is supported
sgl-project/sglang#357724[Bug] Qwen3-VL vision features diverge from Transformers/vLLM in v0.5.17 on fine-grained grounding
sgl-project/sglang#471034[Perf] scheduler: Add opt-in stop-string tail reuse and prefix pruning
sgl-project/sglang#444354[Feat] [NPU] multimodal: Add opt-in cross-request ViT image batching
sgl-project/sglang#347404[Bug] SGLANG_SIMULATE_ACC_LEN silently degrades detokenization to O(nยฒ) โ€” `predict.fill_(100)` emits a byte-fallback token, and the `endswith("\ufffd")` commit gate then never advances the incremental-detokenization offsets
huggingface/lerobot#51064fix(pi0): mask padded actions in the loss
NVIDIA-NeMo/Automodel#44344feat(gemma4_unified): model-owned context parallelism
NVIDIA-NeMo/Automodel#45714fix(qwen): port HF packed fallback boundary guards
NVIDIA-NeMo/Automodel#45834fix(deepseek): add fixed-order TileLang sparse attention backward
NVIDIA-NeMo/Automodel#44814fix(attention): expose backend selection and document precedence
NVIDIA-NeMo/Automodel#44514fix(qwen3.8): restore pretrained generation config
huggingface/accelerate#43304Resuming with load_state + skip_first_batches (checkpointing example) does not restore the shuffle order
vllm-project/llm-compressor#34084Share NVFP4 global scales across Qwen3.5 gated delta net projections
embeddings-benchmark/mteb#58224Add Etsu-Embedding-0.8B-Memory
unslothai/unsloth-zoo#23244train_on_responses_only: find fully masked rows from the Arrow labels column
unslothai/unsloth-zoo#23104Tests: offline merge checks reset hub 0.x sessions so earlier online tests cannot leak
unslothai/unsloth-zoo#21954gpt-oss: preserve left-padding masks in training
ggml-org/llama.cpp#306703model: add Beam (Reflection)
ggml-org/llama.cpp#314713mtmd : add Nemotron 3.5 Transcribe support
ggml-org/llama.cpp#313433gguf-py : fix crash when extra_special_tokens is not a dict
vllm-project/vllm#480763[Performance]: Triton attention drops long-context decode from 3D split-KV to 2D at batch โ‰ฅ ~12, ~doubling ITL
vllm-project/vllm#490313[Bug]: FlashInfer B12x W4A16 runtime repacking retains original and packed MoE weights, causing startup OOM
vllm-project/vllm#505573[Bug]: MiniMax-M3-NVFP4 illegal memory access
vllm-project/vllm#504333[Performance][MooncakeStoreConnector] Overlapping async cold-prefix loads cause duplicate Gets into distinct GPU blocks
vllm-project/vllm#466113[Bug]: No available shared memory broadcast block found in 60 seconds. This typically happens when some processes are hanging or doing some time-consuming work (e.g. compilation, weight/kv cache quantization).
vllm-project/vllm#641193[Fast Start] Support ModelOpt CUTLASS block FP8 cache restarts
vllm-project/vllm#592063[Transformers backend] Migrate ModernBERT embedding, reranker and token classification models
vllm-project/vllm#613203Log EAGLE draft rope base at startup; warn on silent default
vllm-project/vllm#639563[Perf][DSA] Token-sharded sparse-MLA prefill under TP; remove indexer row sharding
vllm-project/vllm#583793[Bugfix][Multimodal] Reject empty Qwen2-VL video samples
vllm-project/vllm#648823[Bug]: Multimodal video loaders ignore --media-io-kwargs num_frames
vllm-project/vllm#645363[Feature][Structured Output] Support continued assistant prefixes
vllm-project/vllm#621603[Bugfix][KV Offload] Deduplicate replicated GQA heads in canonical storage
unslothai/unsloth#162463Tests: gpt-oss dense-expert reload fixture drops replaced experts' quant state
unslothai/unsloth#161783[Bug] Studio: image generation returns 404 after a managed account loads a cached gated model
unslothai/unsloth#163103Studio diffusion: change LoRA weights without reloading the adapter files
unslothai/unsloth#161863[Bug] [Critical Regression / Severity: Incident-level bug] Isolated install via UNSLOTH_STUDIO_HOME silently overwrites existing baseline venv
unslothai/unsloth#163733Notebooks CI: refresh the Colab oracle snapshots to the current GPU runtime
hiyouga/LlamaFactory#108753[export] keep exported tokenizer_config loadable with transformers<5
deepspeedai/DeepSpeed#89803[Hybrid Engine] Require an explicit accelerator opt-in for graph capture
deepspeedai/DeepSpeed#89393feat(rollout): non-greedy sampling on hybrid engine graph decode path
deepspeedai/DeepSpeed#89333[Hybrid Engine] Enable graph capture for HybridEngine on non-CUDA accelerators
deepspeedai/DeepSpeed#86593Add deepspeed.models.glm_moe_dsa: TileLang sparse attention for GLM-5.2
deepspeedai/DeepSpeed#87393Overlap AutoEP's expert weight gradients with the DeepEP backward combine
deepspeedai/DeepSpeed#89593Honor explicit dtype in ZeRO tensor constructors
sgl-project/sglang#473673[Feat] [NPU] attention: support mixed-chunk on the ascend attention backend
sgl-project/sglang#471293[Perf] [NVIDIA] muse-spark: Add B200 runtime optimizations
sgl-project/sglang#477693[Feat] multimodal: Add bounded GPU feature publication admission
sgl-project/sglang#377533[Do not merge - need to clean] GLM-5.3-Flash (glm5_next) LoRA serving + glm53next port rebased
sgl-project/sglang#414033[Benchmark] Add synthetic audio input to bench_serving
sgl-project/sglang#464833[Perf] [NPU] attention: Avoid CPU sequence length synchronization for Kimi K3
sgl-project/sglang#448133[Fix] deepseek_v4: route unquantized BF16/F16/F32 checkpoints off the MXFP4 expert path
sgl-project/sglang#371833AMD MI308X SGLang GLM-5.3-Flash ValueError: The checkpoint you are trying to load has model type `glm5_next` but Transformers does not recognize this architecture.
huggingface/diffusers#151243Quantizer configs are mutated in place, leaking modules_to_not_convert across loads
huggingface/diffusers#151443Fix/flux lora quantization compute embeddings
huggingface/diffusers#151483[core] Enforce minimum supported versions in availability checks
huggingface/diffusers#151473Fix replicated parameter dtypes during tensor-parallel loading
huggingface/lerobot#51553fix(policies): load FAST action tokenizers saved by transformers 4
huggingface/lerobot#47783feat(policies): load GR00T fine-tunes without NVIDIA's base model
huggingface/lerobot#51523perf(pi05): share one read-only prefix cache across the denoise steps
huggingface/lerobot#51513feat(fastwam): add Streaming-WAM to Fast-WAM (2.9x faster warm inference on L40S, no pause at chunk refresh)
huggingface/peft#40503fix: report global dims of tensor-parallel LoRA adapters
huggingface/peft#40673Error message for unmatched target_modules does not reveal that patterns are matched with re.fullmatch
huggingface/peft#38573PeftModelForQuestionAnswering silently drops `token_type_ids` and `position_ids`
NVIDIA-NeMo/Automodel#45753feat(glm): FP8 SFT recipe for GLM-5.3 with model-owned lm_head/indexer precision
NVIDIA-NeMo/Automodel#45853fix(perf): recognize B300 and GB300 for MFU
NVIDIA-NeMo/Automodel#45433feat(deepseek_v41): part-by-part FP8/FP4 base-model loading
NVIDIA-NeMo/Automodel#45763fix(glm5_next): restore pretrained generation config
NVIDIA-NeMo/Automodel#45653feat(deepseek-v41): add packed vision CP and 128K recipe
modelscope/ms-swift#83363ไฝฟ็”จLoraๅพฎ่ฐƒๆ—ถ๏ผŒๆฃ€ๆŸฅ็‚นไฟๅญ˜ๆ—ถๅ‡บ็Žฐadapter name ไธๅŒน้…
NVIDIA/TensorRT-LLM#209733[None][perf] Move vision pixel normalization from CPU to GPU
NVIDIA/TensorRT-LLM#209073[TRTLLM-16460][feat] agent-flow: unified trtllm_modeling_v2 workflow for gated ModelingV2 onboarding and perf tuning
NVIDIA/TensorRT-LLM#207143[TRTLLM-17238][feat] Add PrimTS sparse core backend for Qwen3.8 QSA attention
axolotl-ai-cloud/axolotl#42303feat(kernels): Mistral Large 4 fused MLA attention LoRA + block-FP8 training on transformers 5.22
axolotl-ai-cloud/axolotl#42273fix: follow kernels 0.18's per-module kernelized function copies
axolotl-ai-cloud/axolotl#42313fix: kernelize device for kernels 0.18, tree-recurrence fused Mamba2 on odd row widths
axolotl-ai-cloud/axolotl#42323feat(kernels): model-agnostic grouped-expert kernels (block-FP8 / MXFP8 / Marlin dX) + moe_grouped_mode auto
ml-explore/mlx-lm#20183Add Mistral Large 4 text model
intel/auto-round#25643fix: support nested Qwen2.5-Omni thinker layout
unslothai/unsloth-zoo#23583GLM5 next sparse flex grads test: scale atol with each parameter gradient
unslothai/unsloth-zoo#23573MoE expert LoRA vanilla load test: run on transformers 4.x, and mark the legacy Qwen3-VL-MoE packing it loads silently wrong
unslothai/unsloth-zoo#23563MoE expert LoRA vanilla load test: skip the two checks PEFT < 0.20 fails without Unsloth
unslothai/unsloth-zoo#23543GLM5 next tests: keep assert_close mismatch report in the gradient checks
unslothai/unsloth-zoo#23403rand4x cache tag: import set_torch_config absolutely
unslothai/unsloth-zoo#23383GLM-5.3-Flash flex: keep eager calls with an empty real-query selection dense
ggml-org/llama.cpp#314562gptneox: FFN uses tanh GELU (LLM_FFN_GELU) but GPT-NeoX/Pythia are trained with exact erf GELU
vllm-project/vllm#633692[Multimodal] Add TorchCodec image decoding backend
vllm-project/vllm#590582[Bugfix][MoRIIO] Preserve WRITE receive mappings after consumer abort
vllm-project/vllm#644122[Bugfix][Frontend] Pool HF fast tokenizers in all renderers
vllm-project/vllm#645432[Bugfix][Model] Fix dots_ocr vision tower QKV split for TP > 1
vllm-project/vllm#623952[Feat][2/2] Add Transformers TPSP dense baseline support
vllm-project/vllm#475282[Bug]: DeepSeek-V4-Pro (DeepseekV4ForCausalLM, scale_fmt=ue8m0 / FP4 MoE) produces garbled / degenerate output under tensor parallelism (TP), while data parallelism + expert parallelism (DP+EP) works correctly
vllm-project/vllm#627522[Performance]: Avoidable preemption at hybrid Attention/Mamba decode block transitions
vllm-project/vllm#618882[Distributed] Opt-in FP16 partial sums for tensor-parallel row-parallel layers
vllm-project/vllm#636362[Bugfix][Transformers] Infer mm encoder dynamic_arg_dims from forward signature
vllm-project/vllm#650212[Doc][ROCm] Add verified MI300X and MI325X entries to the Qwen3.8-27Bโ€ฆ
vllm-project/vllm#606082[Bug]: Exllama GPTQ and RDNA3 W4A16 kernels silently produce wrong output when `group_size % 32 != 0` (reproduced on CUDA sm_120 and ROCm gfx1100)
vllm-project/vllm#561802Failed: Cuda error /workspace/csrc/custom_all_reduce.cuh:164 'invalid argument'
vllm-project/vllm#642392[Bug]: Timed-out overflow dequeue returns the wrong message
vllm-project/vllm#649682[WIP][Feature] Qwen GDN BI + MTP + prefix cache integration
vllm-project/vllm#613532[Bugfix][Spec Decode] Fix DFlash capture on FLEX_ATTENTION and honor drafter enforce_eager
vllm-project/vllm#473262[Bug]: Incoherent output depending on tensor-parallel-size during inference on Leanstral (Mistral Small 4 family)
vllm-project/vllm#648992[Bugfix][Multimodal] Honour explicit num_frames in Qwen-VL video loaders
unslothai/unsloth#163402CLI: inference and chat cannot take an image, so a VLM can be trained but not used
unslothai/unsloth#163472CLI: inference and chat take images for vision models
unslothai/unsloth#162122Studio: check the weight index of a module path that only reads as the root
unslothai/unsloth#162422Studio: LFM3-Omni audio input, export, LoRA defaults and Stop for custom-code generate
deepspeedai/DeepSpeed#85022Restore full training state in checkpoint writer engines
deepspeedai/DeepSpeed#87452AutoEP: opt-in skip of the recomputed DeepEP combine
sgl-project/sglang#478302[Bug] Hybrid Mamba model + EAGLE: KV-cache sizing error says to raise --mem-fraction-static when the Mamba state cache is what fills memory
sgl-project/sglang#477702[Perf] DFlash: Use CUDA graphs for draft and idle verify under DP attention
sgl-project/sglang#477292[Feat] qwen3.5: Support native sequence classification
sgl-project/sglang#425512[Feature] Support API keys from secret environment variables
sgl-project/sglang#425502[Fix] Reject removal of an external corpus while its load is pending
sgl-project/sglang#425492[Fix] Preserve Gemma4 null and exponent tool argument types
sgl-project/sglang#373492[Guide] Setup and Troubleshooting for SGLang on V100
sgl-project/sglang#431022[Feat] [NPU] qwen4: Support Qwen3.8-Flash-Next W8A8 inference with QSA on Ascend A2
sgl-project/sglang#403562[Fix] [Ascend] bench_serving: Correct Kimi K2.5 image token counts
sgl-project/sglang#474162[Fix] metrics: Correct MFU/MBU bandwidth overcounting and GraniteMoE AutoTokenizer fallback
sgl-project/sglang#370522[Bug] Qwen3.8-Flash-Next + NEXTN full decode graph: repeated dual-rank invalid-probability asserts on GB10
sgl-project/sglang#370972[Bug] Pretokenized image requests can crash GLM MRoPE with a stale retokenized mask
huggingface/diffusers#151612DreamLite: disabling attention slicing replaces the GQA processor and breaks inference
huggingface/lerobot#50992[Discussion] Has the bf16 update-erasure paper been evaluated or reproduced in LeRobot?
verl-project/verl#84142[model] perf: build Qwen3-VL/Qwen3.5 pos-embed interpolation tensors without Python lists
huggingface/peft#40862Fix broken examples on transformers 5.x (target_modules, dataset IDs, warmup_ratio)
huggingface/peft#39382FIX Reshard on disable_adapter() exit only when needed
huggingface/trl#80062Fail CI test jobs on self-hosted runners when pytest does not complete
huggingface/trl#80582[WIP] Quantize the tiny DeepSeek-V3 and GPT-OSS models like their references (FP8, MXFP4)
huggingface/trl#80502`parse_response` puts Qwen3.5 tool calls into `reasoning_content` when the model skips the prefilled `<think>`
huggingface/trl#79652[BUG] GRPO/RLOO continuous batching casts the trained weights to bf16 in place
NVIDIA-NeMo/Automodel#45822fix(minimax_m3_vl): capture DSpark last-layer features before the final norm
NVIDIA-NeMo/Automodel#45842fix(minimax_m3_vl): accept SM103 (B300/GB300) for the MSA kernels
NVIDIA-NeMo/Automodel#45322feat(moe): support Dist-MoE expert backend
NVIDIA-NeMo/Automodel#45512feat(data): add a common finite-stream loader contract
modelscope/ms-swift#97262fix: Call destroy_process_group() at the end of the trainer pipelines
modelscope/ms-swift#78752่‡ชๅทฑๅ†™็š„ptๅ’ŒvllmๆŽจ็†ๅทฎๅผ‚ๅทจๅคง ๅŽŸๆก†ๆžถ็š„ptๅ’ŒvllmไนŸๆœ‰ๅทฎๅผ‚
modelscope/ms-swift#81642Bidirectional mask for embedding training w/ megatron
modelscope/ms-swift#103072autoawq ๅ’Œgptqmodelไพ่ต–ไธŽtorch็‰ˆๆœฌๅ†ฒ็ช
modelscope/ms-swift#83292ms-swift4.0.1+transformer 5.2.0ๅพฎ่ฐƒqwen3-vl-30b-a3b๏ผŒ้ƒจ็ฝฒ้‡ๅŒ–ๆจกๅž‹ๆ—ถIndexError: Dimension out of range (expected to be in range of [-1, 0], but got 1)
EleutherAI/lm-evaluation-harness#44402Fix xquad, coqa, record and other QA tasks on transformers 5.20 (squad_metrics was removed)
axolotl-ai-cloud/axolotl#41192chat_template: text-diff turn locator trains the role-header token when content starts with whitespace (Phi-3.5, Phi-4)
huggingface/accelerate#43592`gather_for_metrics` keeps padding duplicates for a sized `IterableDataset` with `dispatch_batches=False`
THUDM/slime#18102[Question] ๅคš่ฝฎsft/ๅคš่ฝฎfunction call sftๆ•ฐๆฎ้›†็ป„็ป‡ๅฝขๅผ
vllm-project/llm-compressor#34202add GLM-MoE-DSA MTP support
allenai/open-instruct#21412Add named MILES candidate sources and verifier qualification
ModelCloud/GPTQModel#34802fix(cpu): new cpu kernels regressed trainable api calls from peft
unslothai/unsloth-zoo#22782Merge: apply a trust_remote_code model's registered checkpoint renames, and refuse a merge that places no LoRA
vllm-project/vllm#650751[Bug]: Bi-encoder scoring token limits fail with MistralTokenizer
vllm-project/vllm#650741[Bug]: MCP tool discovery drops tools after the first tools/list page
vllm-project/vllm#650811[CI Failure]: pre-commit mypy on main - process_weights_after_loading redefined in vllm/models/muse_spark/amd/model.py
vllm-project/vllm#650711[Bug]: CustomDataset converts numeric-string prompts to integers before tokenization
vllm-project/vllm#650721[Bug]: ASRDataset uses channel count as duration when filtering stereo audio bytes
vllm-project/vllm#639061[Bug]: FunctionGemma converts escape-delimited string literals into numbers, booleans and null
vllm-project/vllm#650481[CI Failure]: MI300 MoE Kernels Multi-GPU - MoriEPv2 combine block_num 304 exceeds xdb_flag_slots 256 in test_modular_kernel_combinations.py
vllm-project/vllm#507921[Bug]: OverflowError crash in `prompt_logprobs` decode path on PD-disaggregated (external KV connector) setup
vllm-project/vllm#650351[Bug]: Anthropic Messages stream reports the whole prompt as input_tokens in message_start
vllm-project/vllm#634871[Bugfix] Propagate EngineDeadError on AsyncLLM output_handler cancellation
vllm-project/vllm#650151[Bugfix][Tool Parser] Keep granite4 streamed text between tool calls
vllm-project/vllm#650021[Model] Add NVIDIA Nemotron-3-Diarization (offline speaker diarization)
vllm-project/vllm#648091[Bug]: `raw_logprobs` are computed after the structured-output grammar mask
vllm-project/vllm#649561[Bug]: kv_cache_dtype="turboquant_3bit_nc" crashes on Phi-2 (head_size=80) โ€” supports_head_size() claims "any positive value", but the Hadamard rotation is built at the next power of two (128ร—128) and matmul'd with an 80-dim key
vllm-project/vllm#649611[Bugfix][Frontend] Convert Responses content parts to chat parts
vllm-project/vllm#649531[Bug]: `seed=-1` passes config validation and kills engine init inside `np.random.seed` ("Seed must be between 0 and 2**32 - 1") โ€” while `SamplingParams` already treats `-1` as "no seed"
vllm-project/vllm#648851[Bugfix][Pooling] Enable Matryoshka dimensions for EmbeddingGemma 2
vllm-project/vllm#649031[Bugfix][MoE] Avoid double router weighting in TRTLLM FP8 experts
vllm-project/vllm#648451[Bug]: EmbeddingGemma 2 rejects `dimensions` by default although the model supports Matryoshka (128/256/512/768)
unslothai/unsloth#164071UNSLOTH_PARALLEL: one flag and one root mesh for dp / tp / pp / cp, pipeline parallel from the environment, and a capability registry
unslothai/unsloth#164161Packed mRoPE attention: no graph break on warm packed rows
unslothai/unsloth#161691[Bug] Sandbox error that ejects the model?
unslothai/unsloth#164061Pipeline parallel: carry the Qwen3-MoE router auxiliary loss across stages
unslothai/unsloth#164101Block compile: install the router_logits capture hooks before the first block compiles
unslothai/unsloth#163311CLI: no way to pass trust_remote_code, so models with custom code cannot load
unslothai/unsloth#163321CLI: let train load models that need trust_remote_code
unslothai/unsloth#162211[Bug] ShareGPT image lists fail after passing dataset format checks
unslothai/unsloth#162941[Bug] Studio video: MiniMax-H3 on Auto attention crashes with "No available kernel" on Windows (torch without flash attention)
unslothai/unsloth#156321Block Swap blocker: Qwen3.8-Flash-Next-FP8 can't train with offload_layers, its FP8 n-gram table loads without its weight_scale
unslothai/unsloth#162521Tests: compare the disabled 4-bit fused LoRA forward with dequantize + linear, not bitsandbytes' fused GEMM
unslothai/unsloth#161041Saving: write the source tokenizer_class instead of TokenizersBackend
unslothai/unsloth#161021Saving: drop the training truncation and padding from saved tokenizers
hiyouga/LlamaFactory#108681enable fp16 in dtype inference on Intel XPU
deepspeedai/DeepSpeed#89501Fix torch-latest workflow unit failures
deepspeedai/DeepSpeed#88511[AutoTP] Support frozen AutoTP models and manage AutoTP mode per model
deepspeedai/DeepSpeed#88431Keep the parameters a model names for fp32 in fp32 under ZeRO-3
sgl-project/sglang#463031[Perf] diffusion: fix 6 quantization-layer overheads (NVFP4 scale conversion, W8A16, BnB, Nunchaku)
sgl-project/sglang#462481[Feat] hicache: Add unified Engram store and support Qwen4-Exp
sgl-project/sglang#478411[Fix] transcription: Serve /v1/audio/transcriptions for Qwen3-Omni and Gemma 4
sgl-project/sglang#457981[Fix] llava: Load LLaVA-1.5 CLIP tower
sgl-project/sglang#399721[Model] Add BerryLM
sgl-project/sglang#376481[Bug][ROCm] GLM-5.3-MXFP4 TP8 + EAGLE GPU memory fault under high-concurrency long-context load
sgl-project/sglang#376461[Bug] --enable-flashinfer-pure-allreduce is not available on 0.5.18
sgl-project/sglang#376451[Bug] glm45 reasoning parser returns the whole answer as reasoning_content (content empty) when the model skips thinking and stops on EOS
sgl-project/sglang#477861[Feat] multimodal: Share CPU preprocessing artifacts across tokenizer workers
sgl-project/sglang#470601[Fix] observability: Snapshot STANDALONE metrics before settlement
sgl-project/sglang#477461[Fix] diffusion: SenseNova-U1: keep think tokens in attention, fix 4.x import and vision_model init
sgl-project/sglang#476491[Fix] scoring: Return token_logprobs for setwise CausalLM scoring
sgl-project/sglang#475921[CI] ci: Preload `sglang.test.test_utils` in CPU fork workers
sgl-project/sglang#372831[Bug] A non-object entry in a tool-call array discards the valid calls beside it
sgl-project/sglang#372151[Bug] --dp 8 intermittently fails with TCPStore EADDRINUSE on single-node 8ร—H800
sgl-project/sglang#371861[Bug] Qwen3.5 visual encoder weights dropped by PP layer-range filter in load_weights
huggingface/diffusers#151211Fix several quantizer config bugs (AutoRound override, ModelOpt JSON crash, single-file pre_quantized, config mutation, Nunchaku group_size, dataclass eq/repr)
huggingface/diffusers#151631Flux and LTX pipelines do not forward the generator to stochastic scheduler steps
huggingface/lerobot#47751feat(policies): load pi0, pi0.5 and pi0-FAST checkpoints without first building random weights
huggingface/lerobot#51461fix(eo1): give the slow Qwen2.5-VL image processor CPU images
huggingface/lerobot#51621fix(policies): let EO-1, WALL-X and Wall-OSS 0.5 pass torch.export
huggingface/lerobot#51341feat(policies): add Cosmos3
verl-project/verl#84041[hardware, megatron, doc] feat: upgrade Ascend 950 to Megatron-Core 0.18 and add MXFP8 GRPO example
huggingface/peft#40961Update FSDP docs
huggingface/peft#40941reduce_intruder_dimension gives wrong LoRA factors for fan_in_fan_out (Conv1D) and Embedding layers
huggingface/peft#40901layers_to_transform without layers_pattern matches nothing when the layer list is at the model root
huggingface/peft#40881[v1 #4071] Disjoint auxiliary trainable tokens activate only the first adapter
huggingface/peft#40931CI: Bump the third-party-actions group with 5 updates
huggingface/trl#80401Add experimental DecisionTrainer
huggingface/trl#80541Allow `<tool_call>` to close reasoning region in `qwen3_5_template`
huggingface/trl#80521SFTTrainer silently skips ModernBERT Decoder's prediction head
NVIDIA-NeMo/Automodel#45151feat(speculative): add MTP accept-length eval and finetune example
NVIDIA-NeMo/Automodel#45551feat(data): integrate Zephon image training across VLM families
NVIDIA-NeMo/Automodel#45561test(zephon): cover distributed replay and model parity
NVIDIA-NeMo/Automodel#45541refactor(data): enable sharing VLM preprocessing across data load
NVIDIA-NeMo/Automodel#45531feat(data): add Zephon instruction tuning
NVIDIA-NeMo/Automodel#45521feat: support Zephon for text pretraining
NVIDIA-NeMo/Automodel#45501fix(packing): preserve native packed attention through PP
NVIDIA-NeMo/Automodel#45491fix(pp): preserve native model media ownership
NVIDIA-NeMo/Automodel#45481feat(data): add shared packed attention and media handoff
EleutherAI/lm-evaluation-harness#43691Fix greedy token matching in the SGLang generate API
huggingface/accelerate#43191Fix disabled ZeRO-3 initialization contexts
huggingface/accelerate#44751FSDP2 cpu_ram_efficient_loading: honor ignored modules and load buffers
linkedin/Liger-Kernel#15791[Feature] Support GLM-5.3-Flash (glm5_next): fused linear cross-entropy and RMSNorm
linkedin/Liger-Kernel#14471[WIP] Add Qwen4Exp text model support
Blaizzy/mlx-vlm#26401Make the LFM3-VL processor follow the transformers processor contract
Blaizzy/mlx-vlm#26391LFM3-VL processor does not follow the transformers processor contract
Blaizzy/mlx-vlm#2638120261031 -- Diagnostic update: 45-model sweep on 0.7.11 @ 4693cd0e: no crashes, nothing new; LFM3-VL and Qwen3.6 run cleanly
EvolvingLMMs-Lab/lmms-eval#16251[Task] RoboAfford-Eval
vllm-project/llm-compressor#34271[Bug]: NVFP4 MoE example (qwen_30b_a3b.py) gives unstable, damaging expert activation scales: unseeded 20-sample calibration
allenai/open-instruct#21081fix: restore FP32 RoPE frequencies after DeepSpeed BF16 initialization
huggingface/optimum#24551fix #2432: support transformers>=5.0.0 and fix torch.load security warning
unslothai/unsloth-zoo#23601Gemma-4 forced float32: keep norms float32 on full finetuning so lm_head gets float32
unslothai/unsloth-zoo#23521Compiler: keep softmax outputs float32 when the source asks for float32
unslothai/unsloth-zoo#23501Granite-4 hybrid: cast float32-weight norm outputs back to 16-bit without autocast
unslothai/unsloth-zoo#23431Save Llama-4 and Qwen3-VL-MoE expert LoRA in the layout vanilla PEFT loads
unslothai/unsloth-zoo#23481prepare_model_for_training: cast float32 bias masters at every owner of a tied bias
unslothai/unsloth-zoo#23451MoE LoRA stash probe: measure the experts forward, not the gpt-oss decode route
unslothai/unsloth-zoo#23441Qwen3-MoE forced float32: fix full finetuning on GPUs without bfloat16
unslothai/unsloth-zoo#23231Answer the transposed-view check together with the home grouped_mm probe
ggml-org/llama.cpp#309220mtmd : resize Gemma 4 images without letterbox padding
vllm-project/vllm#650730[Bug]: Duplicate MCP server names advertise tools from a different execution server
vllm-project/vllm#633000[Rust Frontend] Complete partial cached HF snapshots when online
vllm-project/vllm#650550[Performance]: MoE top-k gating PDL launch (#59737) slows CUDA graph replay on pre-Hopper GPUs
vllm-project/vllm#639830[Bug]: FunctionGemma tool parser drops every non-string argument (numbers, booleans, arrays, objects)
vllm-project/vllm#645410[Bug]: dots.ocr produces identical "blurry image" output for every region with tensor-parallel size 2
vllm-project/vllm#644180[RFC]: Load plan โ€” an engine-neutral API for sharded weight transfer
vllm-project/vllm#649540[Bug]: `--attention-backend TORCH_SDPA` passes config validation, then dies with "Backend TORCH_SDPA must be registered before use" โ€” a ViT-only tag offered as a valid option, whose `ValueError` bypasses the "Selected backend ... is not valid" path
vllm-project/vllm#624700[Bugfix][LoRA] Scale embedding LoRA deltas by the base layer's embed_scale
unslothai/unsloth#164130Studio: keep a generated chat title in the conversation's language
unslothai/unsloth#163770Studio: apply the mapped chat template once per loaded tokenizer
unslothai/unsloth#157650Studio: show measured transcription progress
unslothai/unsloth#162790Ascend: opt-in eager bootstrap for Torch 2.6/Triton 3.5 on ARM64
unslothai/unsloth#156330Block Swap blocker: offload_layers never engages on Qwen3.8-Flash-Next, the CPU-kept n-gram table makes the map look like two devices
unslothai/unsloth#151590Block Swap: get_peft_model crashes on MoE models with offload_layers (evicted expert weights read as empty)
hiyouga/LlamaFactory#109290[data] Preserve literal placeholders in user content during template encoding
hiyouga/LlamaFactory#108760[xpu] support profiler, module profiler and env report on Intel GPU
hiyouga/LlamaFactory#109270[chat] pass mm_token_type_ids to generate for mrope models
hiyouga/LlamaFactory#109260[data] avoid mutating Qwen video frame inputs
hiyouga/LlamaFactory#109240feat(data): optionally drop overlong SFT examples
hiyouga/LlamaFactory#109230feat(webui): support private Hub export repositories
hiyouga/LlamaFactory#108490[data] guard Qwen2-VL media tokens against SFT truncation
hiyouga/LlamaFactory#109180[npu] Skip unnecessary gradients in NPU fused MoE backward
hiyouga/LlamaFactory#109220fix: handle a single validation sample after CLI parsing
deepspeedai/DeepSpeed#89320[REQUEST][Hybrid Engine] Enable graph capture for HybridEngine on non-CUDA accelerators
deepspeedai/DeepSpeed#89620Fix AutoTP multi-query GPTBigCode training at TP > 1
deepspeedai/DeepSpeed#89610[BUG] AutoTP multi-query GPTBigCode training fails at TP > 1
sgl-project/sglang#438640[Bug] gemma-4-31B-it: audio request crashes the scheduler; /get_model_info misreports has_audio_understanding: true
sgl-project/sglang#477310[Perf] imports: Keep light modules off HF patches, model configs, and the scheduler
sgl-project/sglang#477100[Bug] SM120: fused-A GEMM eligibility accepts K % 256 but the auto-selected CuTe DSL kernel asserts K % 1024 (hidden_size=3584 crashes)
sgl-project/sglang#478090[Fix] attention: Honor custom masks past the causal tile in Triton extend
sgl-project/sglang#477990[Fix] models: Phi-3-4k ignores its 2047-token sliding window (output breaks past 2k tokens)
sgl-project/sglang#477590[Bug] Video input data still leak into logs at --log-requests-level 0/1
sgl-project/sglang#412740[sglang-miles] Recover disk-delta checkpoints safely across resets and retries
sgl-project/sglang#477010[Feat] activevln: Add differentiable actor and SRT weight export
sgl-project/sglang#474440[Bug] NemotronHConfig rejects transformers 5 layer types `linear_attention` / `full_attention` in config.json
sgl-project/sglang#472230[Fix] constrained: Share one compile among concurrent requests for a grammar key
huggingface/diffusers#151680Bump the actions group across 1 directory with 16 updates
huggingface/diffusers#151670UniPC and DPM-Solver with use_flow_sigmas condition the model on integer timesteps while integrating with float sigmas
huggingface/diffusers#151660[training scripts] `--logit_mean` biases timesteps in the opposite direction to the SD3 paper (positive m oversamples low noise)
huggingface/diffusers#151640Add Kandinsky S2V 1.0 (speech-to-video) as a modular pipeline
huggingface/diffusers#151450flux_lora_quantization example: compute_embeddings.py crashes on current transformers
huggingface/lerobot#51940feat(sim): NVIDIA RoboLab benchmark on Isaac Lab
huggingface/lerobot#51890`lerobot-train` profile on one H200 (ACT, SmolVLA, Diffusion): where step time and memory go
huggingface/lerobot#51880perf(pi0): share one read-only prefix cache across the denoise steps
huggingface/lerobot#50670perf(smolvla): precompute cross-attention expert K/V once per chunk
verl-project/verl#80970[fsdp, model] fix: gather DTensor LM heads for fused kernels
huggingface/peft#40850Fix modules_to_save selection when merging an inactive adapter
huggingface/peft#40890FEAT healthcheck: warn when a randomly initialized task head is not saved
huggingface/trl#79600Tool calling is not supported for some popular model families
huggingface/trl#80560GRPO/RLOO under tensor parallelism: TP ranks sample different completions, and the prompt sampler is sized for every process
huggingface/trl#80510Docs: smol-course links in the SFT, DPO and ORPO trainer pages return 404
huggingface/sentence-transformers#43110[fix] Let MSEEvaluator work without a teacher_model
huggingface/sentence-transformers#43100[fix] Stop RerankingEvaluator crashes on single-document samples and empty sample sets
huggingface/sentence-transformers#43090Fix proportional sampling dropping batches beyond estimated child lengths
huggingface/sentence-transformers#43080Reduce CPU XTR training-scorer memory with streaming token top-k
huggingface/sentence-transformers#43070Support variable-length hard negatives in CrossEncoder ranking losses
huggingface/sentence-transformers#43060[fix] Make SparseAutoEncoder picklable so CSR models work with multi-process encoding and DataLoader workers
NVIDIA-NeMo/Automodel#45780MiMo V2 overlength prefilter estimates image tokens with the HF resize helper, so samples over `max_length` are kept and then never truncated
NVIDIA-NeMo/Automodel#45570`tools/merge_lora.py`: merging an `lm_head` LoRA adapter into a model with `tie_word_embeddings=true` silently changes the input embedding table
modelscope/ms-swift#104920fix(dataset): keep message/tool columns intact when map writes run in spawned workers
modelscope/ms-swift#104910Add GPTQModel-backed AWQ export
modelscope/ms-swift#104890[megatron] Wrap DDP/Megatron-FSDP on the current stream in eager mode
modelscope/ms-swift#104870fix(trainers): avoid gradient accumulation scaling during evaluation
FlagOpen/FlagEmbedding#16220fix: preserve dense pooling in encoder checkpoints
FlagOpen/FlagEmbedding#16210fix: load auto models from paths with trailing separators
FlagOpen/FlagEmbedding#16200fix: require explicit opt-in for embedder remote code
FlagOpen/FlagEmbedding#16190fix: preserve layerwise scores across reranker worker batches
huggingface/accelerate#44480Speed up forward with offloaded weights
bitsandbytes-foundation/bitsandbytes#21160LAMB / LARS never update a parameter whose norm is zero (zero-initialised biases, LoRA B matrices, zero-init residual gains)
ml-explore/mlx-lm#20260perf(gemma4): add per_layer_config, KV-shared logits_to_keep prefill trimming, and MTP/MoE sanitization
ml-explore/mlx-lm#20090fix(gpt_oss): detect transformers [E, H, 2F] layout in sanitize() (fixes #1994)
ml-explore/mlx-lm#20250Support transformers_weights in config.json
ml-explore/mlx-lm#20240Fix hardcoded GELU in "gpt_neox" and "gpt_bigcode"
linkedin/Liger-Kernel#13910Add FP32 accumulation option to fused scaled cross entropy
linkedin/Liger-Kernel#15710CuTe DSL RoPE returns wrong results after the first call when CUTE_DSL_ENABLE_TVM_FFI=1
linkedin/Liger-Kernel#15810fix(cutedsl): use safe low-overhead RoPE launch bindings
linkedin/Liger-Kernel#15870Restore torch cross-entropy state in revert_liger_kernel helpers
linkedin/Liger-Kernel#15740fix(cross_entropy): reduce per-token loss in fp32 and scale reduction="none" gradients in place
linkedin/Liger-Kernel#15830fix(grpo): keep the Triton gradient when delta < 1 + epsilon_high
linkedin/Liger-Kernel#15820fix(grpo): mask the luspo loss elementwise to match TRL
linkedin/Liger-Kernel#15760fix(cutedsl): reuse SM90 scaled CE kernels across token counts
linkedin/Liger-Kernel#15800Add apply_liger_kernel_to_glm5_next (GLM-5.3-Flash): fused linear cross-entropy and RMSNorm
linkedin/Liger-Kernel#15770RoPE auto-dispatch prefers CuTe despite higher warmed host latency on SM90
Blaizzy/mlx-vlm#26230Preserve inference state when resuming interrupted responses
EvolvingLMMs-Lab/lmms-eval#16230fix(cache): include effective model identity in response keys
vllm-project/llm-compressor#34310[Bug]: SpinQuantModifier crashes when transform_config is set and overwrites user mappings
vllm-project/llm-compressor#34290[Bug]: Seven unit-offset RMSNorm classes are missing from the norm calibration registry, so smoothing changes those models' outputs
PrimeIntellect-ai/prime-rl#41440fix(qwen3_5): tie lm_head to input embeddings for tied checkpoints
unslothai/unsloth-zoo#23370train_on_responses_only: jump to the next marker or boundary id with list.index instead of stepping every token
unslothai/unsloth-zoo#23260NPU: avoid BF16 LoRA activation round trips while retaining FP32 weight paths

New monkey patches since last run (0)

Something changed on their side or ours: check whether a transformers change forced it.

@@ -12,6 +12,6 @@

Patched by several repos = missing extension point (kernel hub / attention interface / hook).

symbol#reposrepos
*Attention.forward8FlagEmbedding, TensorRT-LLM, axolotl, LlamaFactory, ms-swift, unsloth, unsloth-zoo, verl
*ForCausalLM.forward6Automodel, axolotl, Liger-Kernel, ms-swift, unsloth, unsloth-zoo
*Model.forward6lmms-eval, axolotl, ms-swift, unsloth, unsloth-zoo, verl
*MLP.forward5TensorRT-LLM, axolotl, DeepSpeed, unsloth, verl
*DecoderLayer.forward4axolotl, LlamaFactory, unsloth, verl
*FlashAttention24GPTQModel, ms-swift, sglang, unsloth-zoo
*Experts3axolotl, Liger-Kernel, unsloth-zoo
*Model._create_and_replace3axolotl, ms-swift, unsloth-zoo
*RMSNorm3axolotl, Liger-Kernel, unsloth
PreTrainedModel.from_pretrained3ColossalAI, ms-swift, unsloth
apply_rotary_pos_emb3axolotl, Liger-Kernel, verl
create_causal_mask3GPTQModel, ms-swift, unsloth-zoo
*Experts.forward2unsloth-zoo, verl
*FlashAttention2.forward2LlamaFactory, unsloth
*MLP2axolotl, Liger-Kernel
*Model._postprocess2ROLL, verl
*Model.load_adapter2ms-swift, unsloth-zoo
*Model.save_pretrained2unsloth, unsloth-zoo
*RMSNorm.forward2Automodel, verl
*SdpaAttention.forward2LlamaFactory, unsloth
*SparseMoeBlock2LlamaFactory, verl
CrossEntropyLoss2axolotl, Liger-Kernel
PreTrainedModel.enable_input_require_grads2ms-swift, unsloth
PreTrainedModel.gradient_checkpointing_enable2unsloth, unsloth-zoo
PreTrainedModel.post_init2sglang, unsloth
_init_weights2GPTQModel, ColossalAI
checkpoint2ms-swift, unsloth-zoo
dynamic_module_utils.check_imports2LlamaFactory, ktransformers
get_model_conversion_mapping2unsloth, unsloth-zoo
nn.CrossEntropyLoss2axolotl, Liger-Kernel
nn.LayerNorm2axolotl, Liger-Kernel
*Attention1GPTQModel
*Attention.ATTN_DONE1ColossalAI
*Attention.CU_SEQLENS1ColossalAI
*Attention.DEFAULT_BLOCK_SIZES1diffusers
*Attention.HALF_INDICES1ColossalAI
*Attention.INNER_RING_GROUP1ColossalAI
*Attention.INTER_RING_GROUP1ColossalAI
*Attention.SP_GROUP1ColossalAI
*Attention.SP_STREAM1ColossalAI

Per repo

-
repothreadspatchesnew patches
ggml-org/llama.cpp400
vllm-project/vllm10060
unslothai/unsloth501320
hiyouga/LlamaFactory1280
deepspeedai/DeepSpeed17110
hpcaitech/ColossalAI0140
sgl-project/sglang7140
huggingface/diffusers1330
huggingface/lerobot1900
huggingface/open-r1000
m-bain/whisperX000
verl-project/verl7480
huggingface/peft1800
huggingface/trl1010
kvcache-ai/ktransformers050
huggingface/sentence-transformers000
NVIDIA-NeMo/Automodel0100
modelscope/ms-swift0530
dottxt-ai/outlines000
NVIDIA/TensorRT-LLM020
flairNLP/flair000
EleutherAI/lm-evaluation-harness000
axolotl-ai-cloud/axolotl0830
FlagOpen/FlagEmbedding020
OpenRLHF/OpenRLHF010
huggingface/accelerate000
THUDM/slime000
bitsandbytes-foundation/bitsandbytes000
ml-explore/mlx-lm000
arcee-ai/mergekit000
linkedin/Liger-Kernel01750
Blaizzy/mlx-vlm000
EvolvingLMMs-Lab/lmms-eval020
urchade/GLiNER000
vllm-project/llm-compressor040
allenai/open-instruct000
huggingface/optimum000
embeddings-benchmark/mteb000
alibaba/ROLL070
PrimeIntellect-ai/prime-rl070
intel/auto-round020
ModelCloud/GPTQModel0150
huggingface/optimum-quanto000
unslothai/unsloth-zoo0670
+
repothreadspatchesnew patches
ggml-org/llama.cpp500
vllm-project/vllm10060
unslothai/unsloth501320
hiyouga/LlamaFactory1280
deepspeedai/DeepSpeed17110
hpcaitech/ColossalAI0140
sgl-project/sglang7140
huggingface/diffusers1330
huggingface/lerobot1900
huggingface/open-r1000
m-bain/whisperX000
verl-project/verl7480
huggingface/peft1800
huggingface/trl1110
kvcache-ai/ktransformers050
huggingface/sentence-transformers600
NVIDIA-NeMo/Automodel31100
modelscope/ms-swift13530
dottxt-ai/outlines000
NVIDIA/TensorRT-LLM620
flairNLP/flair000
EleutherAI/lm-evaluation-harness200
axolotl-ai-cloud/axolotl5830
FlagOpen/FlagEmbedding420
OpenRLHF/OpenRLHF010
huggingface/accelerate500
THUDM/slime100
bitsandbytes-foundation/bitsandbytes100
ml-explore/mlx-lm600
arcee-ai/mergekit000
linkedin/Liger-Kernel121750
Blaizzy/mlx-vlm400
EvolvingLMMs-Lab/lmms-eval220
urchade/GLiNER000
vllm-project/llm-compressor740
allenai/open-instruct200
huggingface/optimum200
embeddings-benchmark/mteb400
alibaba/ROLL070
PrimeIntellect-ai/prime-rl170
intel/auto-round420
ModelCloud/GPTQModel1150
huggingface/optimum-quanto000
unslothai/unsloth-zoo24670
All 662 patches
wherepatch
EvolvingLMMs-Lab/lmms-eval/lmms_eval/models/simple/cambrians_vsc.pyCambrianQwenModel.forward = cambrian_qwen2_forward
EvolvingLMMs-Lab/lmms-eval/lmms_eval/models/simple/cambrians_vsr.pyCambrianQwenModel.forward = cambrian_qwen2_forward
FlagOpen/FlagEmbedding/research/llm_embedder/src/utils/llama_patch.pytransformers.models.llama.modeling_llama.LlamaAttention.forward = forward
FlagOpen/FlagEmbedding/research/llm_embedder/src/utils/llama_patch.pytransformers.models.llama.modeling_llama.LlamaModel._prepare_decoder_attention_mask = (
ModelCloud/GPTQModel/gptqmodel/models/base.pymodeling_utils.check_support_param_buffer_assignment = check_support_param_buffer_assignment
ModelCloud/GPTQModel/gptqmodel/models/definitions/intern_s2_preview.pymodeling_module.create_causal_mask = create_causal_mask_compat
ModelCloud/GPTQModel/gptqmodel/utils/hf.pyPreTrainedModel.__getattr__ = __getattr__
ModelCloud/GPTQModel/gptqmodel/utils/hf.pyPreTrainedModel._gptqmodel_legacy_tied_weights_patch = True
ModelCloud/GPTQModel/gptqmodel/utils/hf.pyPreTrainedModel._gptqmodel_missing_all_tied_weights_patch = True
ModelCloud/GPTQModel/gptqmodel/utils/hf.pyPreTrainedModel.get_expanded_tied_weights_keys = get_expanded_tied_weights_keys
ModelCloud/GPTQModel/gptqmodel/utils/hf.pyPreTrainedModel.get_head_mask = get_head_mask
ModelCloud/GPTQModel/gptqmodel/utils/hf.pyPreTrainedModel.is_parallelizable = False
ModelCloud/GPTQModel/gptqmodel/utils/hf.pymodeling_auto.AutoModelForCausalLM.register = classmethod(patched_register)
ModelCloud/GPTQModel/gptqmodel/utils/hf.pymodeling_llama.LlamaAttention = DeepSeekVLV2CompatLlamaAttention
ModelCloud/GPTQModel/gptqmodel/utils/hf.pymodeling_llama.LlamaFlashAttention2 = DeepSeekVLV2CompatLlamaAttention
ModelCloud/GPTQModel/gptqmodel/utils/hf.pymodeling_llama.LlamaFlashAttention2 = original_llama_attention
ModelCloud/GPTQModel/gptqmodel/utils/hf.pymodeling_utils._init_weights = False
ModelCloud/GPTQModel/gptqmodel/utils/hf.pymodeling_utils._init_weights = original_init_flag
ModelCloud/GPTQModel/gptqmodel/utils/hf.pymodeling_utils.is_flash_attn_2_available = is_flash_attn_2_available
NVIDIA-NeMo/Automodel/nemo_automodel/_transformers/auto_model.pyPreTrainedModel._finalize_model_loading = staticmethod(_compat_finalize)
NVIDIA-NeMo/Automodel/nemo_automodel/_transformers/model_init.pyPreTrainedModel.get_init_context = classmethod(_patched_get_init_context)
NVIDIA-NeMo/Automodel/nemo_automodel/_transformers/utils.pyQwen3ForCausalLM.forward = _patched_forward # type: ignore[method-assign]
NVIDIA-NeMo/Automodel/nemo_automodel/components/models/beam/model.pyBeamForCausalLM.parallelizer = PARALLELIZER
NVIDIA-NeMo/Automodel/nemo_automodel/components/models/common/utils.pyTERMSNorm.forward = _patched_rmsnorm_forward
NVIDIA-NeMo/Automodel/nemo_automodel/components/models/deepseek_v4/model.pyDeepseekV4ForCausalLM.parallelizer = PARALLELIZER
NVIDIA-NeMo/Automodel/nemo_automodel/components/models/deepseek_v41/model.pyDeepseekV41ForCausalLM.parallelizer = PARALLELIZER
NVIDIA-NeMo/Automodel/nemo_automodel/components/models/nemotron_v3/model.pyNemotronHForCausalLM.parallelizer = PARALLELIZER
NVIDIA-NeMo/Automodel/nemo_automodel/components/models/qwen3_5/model.pyQwen3_5ForCausalLM.parallelizer = PARALLELIZER
NVIDIA-NeMo/Automodel/nemo_automodel/shared/transformers_patches.pymodeling_t5.T5LayerNorm = _NativeT5LayerNorm
NVIDIA/TensorRT-LLM/tensorrt_llm/tools/layer_wise_benchmarks/mark_utils.pyGatedMLP.forward = nvtx.annotate("GatedMLP")(GatedMLP.forward)
NVIDIA/TensorRT-LLM/tensorrt_llm/tools/layer_wise_benchmarks/mark_utils.pyKimiKDALinearAttention.forward = nvtx.annotate("KimiKDALinearAttention")(
OpenRLHF/OpenRLHF/openrlhf/utils/deepspeed/deepspeed.pytransformers.modeling_flash_attention_utils.deterministic_g = True
PrimeIntellect-ai/prime-rl/src/prime_rl/inference/patches.pyGlm5NextForCausalLM.process_weights_after_loading = process_weights_after_loading
PrimeIntellect-ai/prime-rl/src/prime_rl/inference/patches.pyLoRAModel.from_local_checkpoint = classmethod(_patched_from_local_checkpoint)
PrimeIntellect-ai/prime-rl/src/prime_rl/inference/patches.pyMiniMaxM2ForCausalLM.hf_to_vllm_mapper = WeightsMapper(
PrimeIntellect-ai/prime-rl/src/prime_rl/inference/vllm/qwen38_weight_loading.pyQwen4ExpNGramEmbedding.load_weights = load_weights
PrimeIntellect-ai/prime-rl/src/prime_rl/trainer/models/gpt_oss/attention.pyGptOssAttention.compute_attention = ulysses_attention
PrimeIntellect-ai/prime-rl/src/prime_rl/trainer/models/layers/ulysses_attn.pyAfmoeFlashAttention._compute_attention = _ulysses_compute_attention
PrimeIntellect-ai/prime-rl/src/prime_rl/trainer/models/layers/ulysses_attn.pyFlashAttention._compute_attention = _ulysses_compute_attention
alibaba/ROLL/roll/models/model_providers.pyAutoModelForCausalLMWithValueHead.post_init = no_set_device_hook_post_init
alibaba/ROLL/roll/third_party/megatron/mtp_patcher.pyGPTModel._original_postprocess = original_gpt_postprocess
alibaba/ROLL/roll/third_party/megatron/mtp_patcher.pyGPTModel._postprocess = patched_gpt_postprocess
alibaba/ROLL/roll/third_party/vllm/gdn_patcher.pyGDNAttentionMetadataBuilder._original_build = original_build
alibaba/ROLL/roll/third_party/vllm/gdn_patcher.pyGDNAttentionMetadataBuilder.build = patched_build
alibaba/ROLL/roll/utils/context_parallel/monkey_patch.pyQwen2Model._update_causal_mask = _update_causal_mask
alibaba/ROLL/roll/utils/context_parallel/monkey_patch.pyQwen2Model._update_causal_mask = old_update_causal_mask
axolotl-ai-cloud/axolotl/src/axolotl/integrations/expert_parallel/experts_fn.pyPreTrainedModel.get_correct_experts_implementation = patched # type: ignore[assignment]
axolotl-ai-cloud/axolotl/src/axolotl/integrations/kernels/libs/scattermoe_lora/experts.pyPreTrainedModel.get_correct_experts_implementation = _patched_get_correct
axolotl-ai-cloud/axolotl/src/axolotl/integrations/kernels/libs/scattermoe_lora/nvfp4_moe_loading.pyPreTrainedModel._initialize_missing_keys = patched
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/models/glm4_moe_lite.pymodeling_glm4_moe_lite.Glm4MoeLiteMLP.forward = LigerSwiGLUMLP.forward
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/models/glm4_moe_lite.pymodeling_glm4_moe_lite.Glm4MoeLiteRMSNorm = LigerRMSNorm
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/models/glm4_moe_lite.pymodeling_glm4_moe_lite.apply_rotary_pos_emb = liger_rotary_pos_emb
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/models/glm4_moe_lite.pymodeling_glm4_moe_lite.apply_rotary_pos_emb_interleave = (
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/models/qwen3_5.pymodeling_qwen3_5.Qwen3_5ForCausalLM.forward = lce_forward
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/models/qwen3_5.pymodeling_qwen3_5.Qwen3_5MLP = _liger_swiglu_mlp_wrapper
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/models/qwen3_5.pymodeling_qwen3_5.Qwen3_5RMSNorm = LigerRMSNormForQwen3_5
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/models/qwen3_5.pymodeling_qwen3_5.Qwen3_5RMSNormGated = FusedRMSNormGated
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/models/qwen3_5.pymodeling_qwen3_5.nn.LayerNorm = LigerLayerNorm
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/models/qwen3_5_moe.pymodeling_mod.Qwen3_5MoeForCausalLM.forward = lce_forward
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/models/qwen3_5_moe.pymodeling_mod.Qwen3_5MoeMLP = _liger_swiglu_mlp_wrapper
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/models/qwen3_5_moe.pymodeling_mod.Qwen3_5MoeRMSNorm = LigerRMSNormForQwen3_5Moe
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/models/qwen3_5_moe.pymodeling_mod.Qwen3_5MoeRMSNormGated = FusedRMSNormGated
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/models/qwen3_5_moe.pymodeling_mod.nn.LayerNorm = LigerLayerNorm
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/plugin.pymodeling_cohere_compass.CohereCompassMLP = LigerSwiGLUMLP
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/plugin.pymodeling_cohere_compass.nn.LayerNorm = LigerLayerNorm
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/plugin.pymodeling_gemma4_unified.Gemma4UnifiedRMSNorm = (
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/plugin.pymodeling_gemma4_unified.Gemma4UnifiedTextMLP = _LigerGemma4UnifiedMLP
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/plugin.pymodeling_gemma4_unified.nn.CrossEntropyLoss = LigerCrossEntropyLoss
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/plugin.pymodeling_gemma4_unified.nn.LayerNorm = LigerLayerNorm
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/plugin.pymodeling_jamba.JambaForCausalLM.forward = jamba_lce_forward
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/plugin.pymodeling_jamba.JambaMLP = LigerSwiGLUMLP
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/plugin.pymodeling_jamba.JambaRMSNorm = LigerRMSNorm
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/plugin.pymodeling_jamba.apply_rotary_pos_emb = liger_rotary_pos_emb
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/plugin.pymodeling_jamba.nn.LayerNorm = LigerLayerNorm
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/plugin.pymodeling_ministral3.Ministral3MLP = LigerSwiGLUMLP
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/plugin.pymodeling_ministral3.Ministral3RMSNorm = LigerRMSNorm
axolotl-ai-cloud/axolotl/src/axolotl/integrations/liger/plugin.pymodeling_ministral3.apply_rotary_pos_emb = liger_rotary_pos_emb
axolotl-ai-cloud/axolotl/src/axolotl/loaders/adapter.pyLoraModel._axolotl_clippable_patched = True
axolotl-ai-cloud/axolotl/src/axolotl/loaders/adapter.pyLoraModel._create_and_replace = _patched
axolotl-ai-cloud/axolotl/src/axolotl/loaders/model.pytransformers.integrations.deepspeed.is_deepspeed_zero3_enabled = lambda: (
axolotl-ai-cloud/axolotl/src/axolotl/loaders/model.pytransformers.modeling_utils.is_deepspeed_zero3_enabled = lambda: True
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/accelerate/fsdp2.pyPreTrainedModel._adjust_tied_keys_with_tied_pointers = (
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/accelerate/fsdp2.pyPreTrainedModel._initialize_missing_keys = _patched_initialize_missing_keys
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/accelerate/fsdp2.pyPreTrainedModel._move_missing_keys_from_meta_to_device = _patched_move_missing_keys
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/attention/flex_attn.pytransformers.integrations.flex_attention.WrappedFlexAttention = WrappedFlexAttention
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/btlm_attn_hijack_flash.pymodeling_btlm.BTLMAttention._attn = flashattn_attn
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/kernelize_fixes.pyPreTrainedModel.set_use_kernels = _ORIG_SET_USE_KERNELS
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/kernelize_fixes.pyPreTrainedModel.set_use_kernels = set_use_kernels
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/llama_attn_hijack_flash.pytransformers.loss.loss_utils.fixed_cross_entropy = fa2_fixed_cross_entropy
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/llama_attn_hijack_xformers.pytransformers.models.llama.modeling_llama.LlamaAttention.forward = xformers_forward
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/loss/chunked.pytransformers.loss.loss_utils.ForCausalLMLoss = for_causal_lm_chunked_loss
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/mistral_attn_hijack_flash.pytransformers.models.mistral.modeling_mistral.CrossEntropyLoss = partial(
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/mixtral/__init__.pyMixtralBlockSparseTop2MLP.forward = mlp_forward
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/falcon_h1/modeling.pyFalconH1DecoderLayer.forward = patched_decoder_forward
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/gemma4/fused_attn.pyGemma4TextAttention.forward = _make_fused_forward(original_forward)
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/gemma4/fused_attn.pyGemma4TextDecoderLayer.__call__ = patched_call
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/gemma4/fused_attn.pyGemma4TextDecoderLayer._axolotl_shared_kv_patched = True
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/gemma4_unified/fused_attn.pyGemma4UnifiedTextAttention._axolotl_fused_attn_patched = True
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/gemma4_unified/fused_attn.pyGemma4UnifiedTextAttention.forward = _make_fused_forward(original_forward)
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/gemma4_unified/fused_attn.pyGemma4UnifiedTextDecoderLayer.__call__ = patched_call
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/gemma4_unified/fused_attn.pyGemma4UnifiedTextDecoderLayer._axolotl_shared_kv_patched = True
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/glm4_moe_lite/fused_attn.pyGlm4MoeLiteAttention._axolotl_fused_attn_patched = True
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/glm4_moe_lite/fused_attn.pyGlm4MoeLiteAttention.forward = _make_fused_forward(Glm4MoeLiteAttention.forward)
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/llama4/modeling.pymodeling_llama4.Llama4TextExperts = Llama4TextExperts
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/llama4/modeling.pymodeling_llama4.Llama4TextExperts = old_lamma_4_text_experts
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/pixtral/modeling_flash_attention_utils.pymodeling_flash_attention_utils._is_packed_sequence = fixed_is_packed_sequence
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/pixtral/modeling_flash_attention_utils.pymodeling_flash_attention_utils._is_packed_sequence = old_fn
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/qwen3/fused_attn.pyQwen3Attention._axolotl_fused_attn_patched = True
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/qwen3/fused_attn.pyQwen3Attention.forward = _make_fused_forward()
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/qwen3_5/fused_attn.pyQwen3_5Attention._axolotl_fused_attn_patched = True
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/qwen3_5/fused_attn.pyQwen3_5Attention.forward = _make_fused_forward()
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/qwen3_5_moe/fused_attn.pyQwen3_5MoeAttention._axolotl_fused_attn_patched = True
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/qwen3_5_moe/fused_attn.pyQwen3_5MoeAttention.forward = _make_fused_forward()
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/qwen3_moe/fused_attn.pyQwen3MoeAttention._axolotl_fused_attn_patched = True
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/qwen3_moe/fused_attn.pyQwen3MoeAttention.forward = _make_fused_forward()
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/qwen3_next/modeling.pyQwen3NextDecoderLayer.forward = original_decoder_forward
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/qwen3_next/modeling.pyQwen3NextDecoderLayer.forward = patched_decoder_forward
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/qwen3_vl/fused_attn.pyQwen3VLTextAttention._axolotl_fused_attn_patched = True
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/models/qwen3_vl/fused_attn.pyQwen3VLTextAttention.forward = _make_fused_forward()
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/moe_quant.pytransformers.core_model_loading.set_param_for_module = _patched_set_param_for_module
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/moe_quant.pytransformers.modeling_utils.caching_allocator_warmup = _noop_warmup
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/multipack.pymodeling_arch._get_unpad_data = get_unpad_data
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/peft/nf4.pyLoraModel.merge_and_unload = merge_and_unload
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/quantized_init.pyPreTrainedModel._initialize_weights = _initialize_weights
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/stablelm_attn_hijack_flash.pymodeling_stablelm.Attention.forward = flashattn_attn
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/stablelm_attn_hijack_flash.pymodeling_stablelm.DecoderLayer.forward = decoder_layer_forward
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/stablelm_attn_hijack_flash.pymodeling_stablelm.StableLMEpochModel.forward = stablelm_model_forward
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/transformers_fa_utils.pytransformers.modeling_flash_attention_utils.fa_peft_integration_check = partial(
axolotl-ai-cloud/axolotl/src/axolotl/monkeypatch/transformers_num_parameters.pyPreTrainedModel.resize_token_embeddings = resize_token_embeddings # type: ignore[method-assign]
deepspeedai/DeepSpeed/deepspeed/inference/engine.pyDeepSpeedSelfAttention.num_layers = 0
deepspeedai/DeepSpeed/deepspeed/linear/context_manager.pytransformers.AutoModelForCausalLM.from_config = from_config
deepspeedai/DeepSpeed/deepspeed/linear/context_manager.pytransformers.AutoModelForCausalLM.from_config = self._orig_causallm_config
deepspeedai/DeepSpeed/deepspeed/linear/context_manager.pytransformers.AutoModelForCausalLM.from_pretrained = from_pretrained
deepspeedai/DeepSpeed/deepspeed/linear/context_manager.pytransformers.AutoModelForCausalLM.from_pretrained = self._orig_causallm_pretrained
deepspeedai/DeepSpeed/deepspeed/ops/transformer/inference/ds_attention.pyDeepSpeedSelfAttention._qkv_buffers = [
deepspeedai/DeepSpeed/deepspeed/ops/transformer/inference/ds_attention.pyDeepSpeedSelfAttention.num_layers = DeepSpeedSelfAttention.num_layers + 1
deepspeedai/DeepSpeed/deepspeed/ops/transformer/inference/ds_mlp.pyDeepSpeedMLP._inter_w_buffers = [
deepspeedai/DeepSpeed/deepspeed/ops/transformer/inference/triton/attention.pyTritonSelfAttention.num_layers = TritonSelfAttention.num_layers + 1
deepspeedai/DeepSpeed/deepspeed/runtime/sequence_parallel/ulysses_sp.pymodeling_llama.LlamaMLP.forward = mlp_forward_new
deepspeedai/DeepSpeed/deepspeed/runtime/sequence_parallel/ulysses_sp.pymodeling_llama.LlamaMLP.forward = tiled_mlp_forward
hiyouga/LlamaFactory/src/llamafactory/extras/misc.pytransformers.dynamic_module_utils.check_imports = get_relative_imports
hiyouga/LlamaFactory/src/llamafactory/model/model_utils/longlora.pyLlamaAttention.forward = llama_attention_forward
hiyouga/LlamaFactory/src/llamafactory/model/model_utils/longlora.pyLlamaFlashAttention2.forward = llama_flash_attention_2_forward
hiyouga/LlamaFactory/src/llamafactory/model/model_utils/longlora.pyLlamaSdpaAttention.forward = llama_sdpa_attention_forward
hiyouga/LlamaFactory/src/llamafactory/model/model_utils/visual.pytransformers.models.llava.modeling_llava.LlavaMultiModalProjector = LlavaMultiModalProjectorForYiVL
hiyouga/LlamaFactory/src/llamafactory/model/patcher.pyQwen3_5DecoderLayer.forward = _patched_decoder_forward
hiyouga/LlamaFactory/src/llamafactory/model/patcher.pyQwen3_5MoeDecoderLayer.forward = _patched_decoder_forward
hiyouga/LlamaFactory/src/llamafactory/model/patcher.pymodeling_qwen3_omni_moe.Qwen3OmniMoeThinkerTextSparseMoeBlock = Qwen3OmniMoeThinkerTextSparseMoeBlock
hpcaitech/ColossalAI/colossalai/lazy/pretrained.pyPreTrainedModel.from_pretrained = classmethod(PretrainedManager.old_from_pretrained)
hpcaitech/ColossalAI/colossalai/lazy/pretrained.pyPreTrainedModel.from_pretrained = new_from_pretrained
hpcaitech/ColossalAI/colossalai/legacy/inference/quant/smoothquant/models/base_model.pytransformers.modeling_utils._init_weights = False
hpcaitech/ColossalAI/colossalai/shardformer/layer/attn.pyColoAttention._flash_kernel_dispatch = ColoAttention._flash_kernel_dispatch.load()
hpcaitech/ColossalAI/colossalai/shardformer/layer/attn.pyColoAttention._flash_kernel_dispatch = FlashAttentionDaoLoader()
hpcaitech/ColossalAI/colossalai/shardformer/layer/attn.pyColoAttention._kernel_dispatch_map = {
hpcaitech/ColossalAI/colossalai/shardformer/layer/attn.pyRingAttention.ATTN_DONE = torch.cuda.Event()
hpcaitech/ColossalAI/colossalai/shardformer/layer/attn.pyRingAttention.CU_SEQLENS = cu_seqlens
hpcaitech/ColossalAI/colossalai/shardformer/layer/attn.pyRingAttention.HALF_INDICES = (half_idx_front, half_idx_back)
hpcaitech/ColossalAI/colossalai/shardformer/layer/attn.pyRingAttention.INNER_RING_GROUP = inner_ring_group
hpcaitech/ColossalAI/colossalai/shardformer/layer/attn.pyRingAttention.INTER_RING_GROUP = inter_ring_group
hpcaitech/ColossalAI/colossalai/shardformer/layer/attn.pyRingAttention.SP_GROUP = sp_group
hpcaitech/ColossalAI/colossalai/shardformer/layer/attn.pyRingAttention.SP_STREAM = torch.cuda.Stream()
hpcaitech/ColossalAI/colossalai/shardformer/layer/attn.pyRingAttention.TOTAL_SEQLEN = b * sq
huggingface/diffusers/examples/research_projects/pytorch_xla/inference/flux/flux_inference.pyFlashAttention.DEFAULT_BLOCK_SIZES = {
huggingface/diffusers/examples/research_projects/pytorch_xla/inference/flux/flux_inference_spmd.pyFlashAttention.DEFAULT_BLOCK_SIZES = {
huggingface/diffusers/src/diffusers/pipelines/prx/__init__.pytransformers.T5GemmaEncoder = T5GemmaEncoder
huggingface/trl/trl/_compat.pytransformers.cache_utils.HybridCache = Cache
intel/auto-round/auto_round/modeling/unfused_moe/__init__.pytransformers.modeling_utils.get_checkpoint_conversion_mapping = get_checkpoint_conversion_mapping_ar
intel/auto-round/auto_round/utils/common.pytransformers.modeling_utils.no_init_weights = no_init_weights
kvcache-ai/ktransformers/archive/kt-sft/ktransformers/operators/dynamic_attention.pyDynamicScaledDotProductAttention.cpu_infer = CPUInfer(threads_num)
kvcache-ai/ktransformers/archive/kt-sft/ktransformers/operators/models.pyKLlamaModel.dynamic_sdpa = DynamicScaledDotProductAttention(
kvcache-ai/ktransformers/archive/kt-sft/ktransformers/sft/metrics_utils/misc.pytransformers.dynamic_module_utils.check_imports = get_relative_imports
kvcache-ai/ktransformers/archive/ktransformers/operators/dynamic_attention.pyDynamicScaledDotProductAttention.cpu_infer = CPUInfer(threads_num)
kvcache-ai/ktransformers/archive/ktransformers/operators/models.pyKLlamaModel.dynamic_sdpa = DynamicScaledDotProductAttention(
linkedin/Liger-Kernel/src/liger_kernel/transformers/llama4_rope.pymodeling_module.apply_rotary_emb = liger_llama4_text_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/llama4_rope.pymodeling_module.vision_apply_rotary_emb = liger_llama4_vision_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pyGemma4TextForCausalLM.forward = causal_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_deepseek_v2.DeepseekV2Experts = LigerExperts
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_deepseek_v2.DeepseekV2ForCausalLM.forward = deepseek_v2_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_deepseek_v2.DeepseekV2MLP = LigerQwen3MoeSwiGLUMLP
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_deepseek_v2.DeepseekV2RMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_deepseek_v3.DeepseekV3Experts = LigerExperts
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_deepseek_v3.DeepseekV3ForCausalLM.forward = deepseek_v3_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_deepseek_v3.DeepseekV3MLP = LigerQwen3MoeSwiGLUMLP
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_deepseek_v3.DeepseekV3RMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_deepseek_v4.DeepseekV4ForCausalLM.forward = deepseek_v4_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_deepseek_v4.DeepseekV4RMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_exaone4.Exaone4ForCausalLM.forward = exaone4_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_exaone4.Exaone4MLP = LigerSwiGLUMLP
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_exaone4.Exaone4RMSNorm = Exaone4LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_exaone4.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_falcon_h1.FalconH1ForCausalLM.forward = falcon_h1_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_falcon_h1.FalconH1MLP = LigerFalconH1SwiGLUMLP
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_falcon_h1.FalconH1RMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_falcon_h1.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_gemma.GemmaForCausalLM.forward = gemma_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_gemma.GemmaMLP = LigerGEGLUMLP
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_gemma.GemmaRMSNorm = LigerRMSNormForGemma
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_gemma.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_gemma2.Gemma2ForCausalLM.forward = gemma2_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_gemma2.Gemma2MLP = LigerGEGLUMLP
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_gemma2.Gemma2RMSNorm = LigerRMSNormForGemma2
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_gemma2.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_gemma3.Gemma3ForCausalLM.forward = causal_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_gemma3.Gemma3ForConditionalGeneration.forward = multimodal_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_gemma3.Gemma3MLP = LigerGEGLUMLP
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_gemma3.Gemma3RMSNorm = LigerRMSNormForGemma3
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_gemma3.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_gemma3.nn.CrossEntropyLoss = LigerCrossEntropyLoss
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_gemma4.Gemma4ForCausalLM.forward = causal_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_gemma4.Gemma4ForConditionalGeneration.forward = multimodal_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_gemma4.Gemma4RMSNorm = LigerRMSNormForGemma4
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_gemma4.Gemma4TextMLP = LigerGEGLUMLPForGemma4
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_glm4.Glm4ForCausalLM.forward = glm4_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_glm4.Glm4MLP = LigerPhi3SwiGLUMLP
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_glm4.Glm4RMSNorm = LigerRMSNormForGlm4
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_glm4v.Glm4vForConditionalGeneration.forward = glm4v_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_glm4v.Glm4vRMSNorm = LigerRMSNormForGlm4
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_glm4v_moe.Glm4vMoeForConditionalGeneration.forward = glm4v_moe_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_glm4v_moe.Glm4vMoeRMSNorm = LigerRMSNormForGlm4
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_glm4v_moe.Glm4vMoeTextNaiveMoe = LigerExperts
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_glm4v_moe.Glm4vMoeTextRMSNorm = LigerRMSNormForGlm4
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_gpt_oss.GptOssForCausalLM.forward = gpt_oss_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_gpt_oss.GptOssRMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_gpt_oss.apply_rotary_pos_emb = liger_rotary_pos_emb_gpt_oss
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_granite.GraniteMLP = LigerSwiGLUMLP
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_granite.GraniteRMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_granite.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_hunyuan_v1_dense.HunYuanDenseV1ForCausalLM.forward = hunyuan_v1_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_hunyuan_v1_dense.HunYuanDenseV1MLP = LigerHunyuanV1SwiGLUMLP
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_hunyuan_v1_dense.HunYuanDenseV1RMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_hunyuan_v1_dense.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_hunyuan_v1_moe.HunYuanMoEV1Experts = LigerExperts
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_hunyuan_v1_moe.HunYuanMoEV1ForCausalLM.forward = hunyuan_v1_moe_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_hunyuan_v1_moe.HunYuanMoEV1RMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_hunyuan_v1_moe.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_internvl.InternVLForConditionalGeneration.forward = internvl_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_internvl.InternVLVisionRMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_internvl.nn.LayerNorm = LigerLayerNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_llama.LlamaForCausalLM.forward = llama_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_llama.LlamaMLP = LigerSwiGLUMLP
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_llama.LlamaRMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_llama.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_llama4.CrossEntropyLoss = LigerCrossEntropyLoss
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_llama4.Llama4ForCausalLM.forward = llama4_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_llama4.Llama4TextMLP = LigerSwiGLUMLP
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_llama4.Llama4TextRMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_llava.LlavaForConditionalGeneration.forward = llava_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_ministral.CrossEntropyLoss = LigerCrossEntropyLoss
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_ministral.MinistralForCausalLM.forward = ministral_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_ministral.MinistralMLP = LigerSwiGLUMLP
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_ministral.MinistralRMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_ministral.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_mistral.CrossEntropyLoss = LigerCrossEntropyLoss
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_mistral.MistralForCausalLM.forward = mistral_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_mistral.MistralMLP = LigerSwiGLUMLP
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_mistral.MistralRMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_mistral.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_mixtral.MixtralExperts = LigerExperts
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_mixtral.MixtralForCausalLM.forward = mixtral_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_mixtral.MixtralRMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_mixtral.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_mllama.MllamaForCausalLM.forward = mllama_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_mllama.MllamaTextMLP = LigerSwiGLUMLP
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_mllama.MllamaTextRMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_mllama.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_mllama.nn.LayerNorm = LigerLayerNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_muse_glimmer.MuseGlimmerForConditionalGeneration.forward = muse_glimmer_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_muse_glimmer.MuseGlimmerRMSNorm = LigerRMSNormForMuseGlimmer
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_muse_glimmer.MuseGlimmerTextCenteredRMSNorm = LigerRMSNormForMuseGlimmerTextCentered
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_muse_glimmer.MuseGlimmerTextMLP = LigerSwiGLUMLPForMuseGlimmer
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_muse_glimmer.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_nemotron.CrossEntropyLoss = LigerCrossEntropyLoss
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_nemotron.NemotronForCausalLM.forward = nemotron_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_olmo2.Olmo2ForCausalLM.forward = olmo2_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_olmo2.Olmo2MLP = LigerSwiGLUMLP
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_olmo2.Olmo2RMSNorm = LigerRMSNormForOlmo2
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_olmo2.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_olmo3.Olmo3ForCausalLM.forward = olmo3_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_olmo3.Olmo3MLP = LigerSwiGLUMLP
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_olmo3.Olmo3RMSNorm = LigerRMSNormForOlmo2 # same as olmo2
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_olmo3.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_paligemma.PaliGemmaForConditionalGeneration.forward = lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_paligemma.nn.CrossEntropyLoss = LigerCrossEntropyLoss
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_phi3.Phi3ForCausalLM.forward = phi3_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_phi3.Phi3MLP = LigerPhi3SwiGLUMLP
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_phi3.Phi3RMSNorm = LigerRMSNorm # Same as Llama
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_phi3.apply_rotary_pos_emb = liger_rotary_pos_emb # Same as Gemma
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_pixtral.PixtralMLP = LigerSwiGLUMLP
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_pixtral.PixtralRMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_pixtral.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen2.Qwen2ForCausalLM.forward = qwen2_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen2.Qwen2MLP = LigerSwiGLUMLP
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen2.Qwen2RMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen2.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen2_5_vl.CrossEntropyLoss = LigerCrossEntropyLoss
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen2_5_vl.Qwen2MLP = LigerSwiGLUMLP
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen2_5_vl.Qwen2RMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen2_5_vl.Qwen2_5_VLForConditionalGeneration.forward = qwen2_5_vl_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen2_5_vl.Qwen2_5_VLRMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen2_5_vl.apply_multimodal_rotary_pos_emb = liger_multimodal_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen2_vl.CrossEntropyLoss = LigerCrossEntropyLoss
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen2_vl.LayerNorm = LigerLayerNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen2_vl.Qwen2MLP = LigerSwiGLUMLP
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen2_vl.Qwen2RMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen2_vl.Qwen2VLForConditionalGeneration.forward = qwen2_vl_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen2_vl.Qwen2VLRMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen2_vl.apply_multimodal_rotary_pos_emb = liger_multimodal_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3.Qwen3ForCausalLM.forward = qwen3_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3.Qwen3MLP = LigerSwiGLUMLP
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3.Qwen3RMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_5.Qwen3_5ForCausalLM.forward = qwen3_5_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_5.Qwen3_5ForConditionalGeneration.forward = qwen3_5_lce_forward_for_multimodal
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_5.Qwen3_5MLP = LigerQwen3MoeSwiGLUMLP
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_5.Qwen3_5RMSNorm = LigerRMSNormForQwen3Next
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_5.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_5.apply_rotary_pos_emb_vision = liger_rotary_pos_emb_vision
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_5_moe.Qwen3_5MoeExperts = LigerExperts
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_5_moe.Qwen3_5MoeForCausalLM.forward = qwen3_5_moe_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_5_moe.Qwen3_5MoeForConditionalGeneration.forward = (
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_5_moe.Qwen3_5MoeRMSNorm = LigerRMSNormForQwen3Next
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_5_moe.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_5_moe.apply_rotary_pos_emb_vision = liger_rotary_pos_emb_vision
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_moe.Qwen3MoeExperts = LigerExperts
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_moe.Qwen3MoeForCausalLM.forward = qwen3_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_moe.Qwen3MoeRMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_moe.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_next.Qwen3NextExperts = LigerExperts
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_next.Qwen3NextForCausalLM.forward = qwen3_next_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_next.Qwen3NextRMSNorm = LigerRMSNormForQwen3Next
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_next.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_vl.Qwen3VLForConditionalGeneration.forward = qwen3_vl_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_vl.Qwen3VLTextRMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_vl.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_vl.apply_rotary_pos_emb_vision = liger_rotary_pos_emb_vision
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_vl_moe.Qwen3VLMoeForConditionalGeneration.forward = qwen3_vl_moe_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_vl_moe.Qwen3VLMoeTextExperts = LigerExperts
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_vl_moe.Qwen3VLMoeTextRMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_vl_moe.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_qwen3_vl_moe.apply_rotary_pos_emb_vision = liger_rotary_pos_emb_vision
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_siglip.nn.LayerNorm = LigerLayerNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_smollm3.SmolLM3ForCausalLM.forward = smollm3_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_smollm3.SmolLM3MLP = LigerSwiGLUMLP
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_smollm3.SmolLM3RMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_smollm3.apply_rotary_pos_emb = liger_rotary_pos_emb
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_smolvlm.SmolVLMForConditionalGeneration.forward = smolvlm_lce_forward
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_smolvlm.SmolVLMRMSNorm = LigerRMSNorm
linkedin/Liger-Kernel/src/liger_kernel/transformers/monkey_patch.pymodeling_smolvlm.nn.LayerNorm = LigerLayerNorm
modelscope/ms-swift/swift/megatron/utils/convert_utils.pyTEDotProductAttention.forward = forward
modelscope/ms-swift/swift/megatron/utils/convert_utils.pyTEDotProductAttention.forward = new_forward
modelscope/ms-swift/swift/model/models/deepseek.pyUnlimitedOCRModel._swift_multi_gpu_patched = True
modelscope/ms-swift/swift/model/models/deepseek.pyUnlimitedOCRModel.forward = _patched_forward
modelscope/ms-swift/swift/model/models/deepseek.pymodeling_module.torch.cat = _orig_cat
modelscope/ms-swift/swift/model/models/deepseek.pymodeling_module.torch.cat = _safe_cat
modelscope/ms-swift/swift/model/models/deepseek.pytransformers.models.llama.modeling_llama.LlamaFlashAttention2 = None
modelscope/ms-swift/swift/model/models/internlm.pyPreTrainedModel._old_enable_input_require_grads = old_enable_input_require_grads
modelscope/ms-swift/swift/model/models/internlm.pyPreTrainedModel.enable_input_require_grads = patched_enable_input_require_grads
modelscope/ms-swift/swift/model/models/llava.pyLlavaLlamaForCausalLM.__old_forward = forward
modelscope/ms-swift/swift/model/models/llava.pyLlavaLlamaForCausalLM.forward = _new_forward
modelscope/ms-swift/swift/model/models/spark.pymodeling_module._swift_patched = True
modelscope/ms-swift/swift/model/models/spark.pysetattr(modeling_module, name, Spark2_5Loader._compat_mask_fn(getattr(modeling_module, name)))
modelscope/ms-swift/swift/model/models/valley.pyValleyQwen2ForCausalLM._origin_forward = forward
modelscope/ms-swift/swift/model/models/valley.pyValleyQwen2ForCausalLM.forward = new_forward
modelscope/ms-swift/swift/model/models/xing4_0.pymodeling_module.Xing4_0HyperConnection.forward = torch.compile(
modelscope/ms-swift/swift/model/models/xing4_0.pymodeling_module.Xing4_0PreTrainedModel._can_set_experts_implementation_cached_value = True
modelscope/ms-swift/swift/model/models/xing4_0.pymodeling_module._swift_patch_options = patch_options
modelscope/ms-swift/swift/model/patcher.pyPreTrainedModel._get_no_split_modules = _get_no_split_modules
modelscope/ms-swift/swift/model/patcher.pyPreTrainedModel._get_no_split_modules = _new_get_no_split_modules
modelscope/ms-swift/swift/model/patcher.pyPreTrainedModel.from_pretrained = _new_from_pretrained
modelscope/ms-swift/swift/model/patcher.pyPreTrainedModel.from_pretrained = classmethod(from_pretrained)
modelscope/ms-swift/swift/model/patcher.pytransformers.modeling_utils.get_balanced_memory = lambda *args, **kwargs: {}
modelscope/ms-swift/swift/model/patcher.pytransformers.modeling_utils.infer_auto_device_map = _infer_auto_device_map_patch
modelscope/ms-swift/swift/sequence_parallel/sequence_parallel.pymodeling_flash_attention_utils._flash_attention_forward = flash_attention_forward
modelscope/ms-swift/swift/sequence_parallel/sequence_parallel.pymodeling_flash_attention_utils._flash_attention_forward_origin = _flash_attention_forward
modelscope/ms-swift/swift/template/base.pymodeling_module._flash_attention_forward = _flash_attention_forward
modelscope/ms-swift/swift/template/base.pymodeling_module._flash_attention_forward = _origin_flash_attention_forward
modelscope/ms-swift/swift/template/templates/glm.pymodeling_module.create_causal_mask = new_create_causal_mask
modelscope/ms-swift/swift/trainers/mixin.pytransformers.modeling_utils.checkpoint = _new_checkpoint
modelscope/ms-swift/swift/tuners/peft.pyAdaLoraModel._create_and_replace = _create_and_replace_hook
modelscope/ms-swift/swift/tuners/peft.pyAdaLoraModel._create_and_replace_origin = AdaLoraModel._create_and_replace
modelscope/ms-swift/swift/tuners/peft.pyAdaLoraModel.forward = adalora_forward
modelscope/ms-swift/swift/tuners/peft.pyBOFTModel._create_and_replace = _create_and_replace_hook
modelscope/ms-swift/swift/tuners/peft.pyBOFTModel._create_and_replace_origin = BOFTModel._create_and_replace
modelscope/ms-swift/swift/tuners/peft.pyBoneModel._create_and_replace = _create_and_replace_hook
modelscope/ms-swift/swift/tuners/peft.pyBoneModel._create_and_replace_origin = BoneModel._create_and_replace
modelscope/ms-swift/swift/tuners/peft.pyFourierFTModel._create_and_replace = _create_and_replace_hook
modelscope/ms-swift/swift/tuners/peft.pyFourierFTModel._create_and_replace_origin = FourierFTModel._create_and_replace
modelscope/ms-swift/swift/tuners/peft.pyLoraModel.__init__ = __new_init__
modelscope/ms-swift/swift/tuners/peft.pyLoraModel.__init_origin__ = LoraModel.__init__
modelscope/ms-swift/swift/tuners/peft.pyLoraModel._cast_adapter_dtype = _cast_adapter_dtype_hook
modelscope/ms-swift/swift/tuners/peft.pyLoraModel._cast_adapter_dtype_origin = LoraModel._cast_adapter_dtype
modelscope/ms-swift/swift/tuners/peft.pyLoraModel._create_and_replace = _create_and_replace_hook
modelscope/ms-swift/swift/tuners/peft.pyLoraModel._create_and_replace_origin = LoraModel._create_and_replace
modelscope/ms-swift/swift/tuners/peft.pyPeftModel.activate_adapter = PeftModel.set_adapter
modelscope/ms-swift/swift/tuners/peft.pyPeftModel.create_optimizer_param_groups = create_optimizer_param_groups
modelscope/ms-swift/swift/tuners/peft.pyPeftModel.deactivate_adapter = partial(dummy_function, func='deactivate_adapter')
modelscope/ms-swift/swift/tuners/peft.pyPeftModel.load_adapter = load_adapter
modelscope/ms-swift/swift/tuners/peft.pyPeftModel.load_adapter_origin = PeftModel.load_adapter
modelscope/ms-swift/swift/tuners/peft.pyPeftModel.set_active_adapters = partial(dummy_function, func='set_active_adapters')
modelscope/ms-swift/swift/tuners/peft.pyVeraModel._create_and_replace = _create_and_replace_hook
modelscope/ms-swift/swift/tuners/peft.pyVeraModel._create_and_replace_origin = VeraModel._create_and_replace
sgl-project/sglang/python/sglang/multimodal_gen/runtime/layers/attention/layer.pyUSPAttention._usp_a2a_stream = torch.get_device_module().Stream()
sgl-project/sglang/python/sglang/multimodal_gen/runtime/loader/utils.pyPreTrainedModel.post_init = lambda self: None
sgl-project/sglang/python/sglang/srt/multimodal/processors/qwen_vl.pytransformers.models.qwen2_vl.image_processing_qwen2_vl_fast.Qwen2VLImageProcessorFast._preprocess = fast_preprocess_cpu
sgl-project/sglang/python/sglang/srt/utils/hf_transformers_patches.pymodeling_llama.LlamaFlashAttention2 = modeling_llama.LlamaAttention
unslothai/unsloth-zoo/unsloth_zoo/compiler.pymodeling_file.__UNSLOTH_PATCHED__ = True
unslothai/unsloth-zoo/unsloth_zoo/compiler.pymodeling_file.__UNSLOTH_SUPPORTS_SDPA__ = False
unslothai/unsloth-zoo/unsloth_zoo/compiler.pymodeling_file.__UNSLOTH_SUPPORTS_SDPA__ = final_supports_sdpa
unslothai/unsloth-zoo/unsloth_zoo/compiler.pymodeling_file.unsloth_mhc_projection = unsloth_mhc_projection
unslothai/unsloth-zoo/unsloth_zoo/compiler.pymodeling_file.unsloth_sinkhorn_knopp = unsloth_sinkhorn_knopp
unslothai/unsloth-zoo/unsloth_zoo/compiler.pysetattr(modeling_file, key, value)
unslothai/unsloth-zoo/unsloth_zoo/empty_model.pyBitsAndBytesModelLoader._stack_quantization_states = (
unslothai/unsloth-zoo/unsloth_zoo/empty_model.pyGemma4ForCausalLM.get_expert_mapping = get_expert_mapping
unslothai/unsloth-zoo/unsloth_zoo/empty_model.pySiglipVisionModel._init_weights = SiglipVisionModel._original_initialize_weights
unslothai/unsloth-zoo/unsloth_zoo/empty_model.pySiglipVisionModel._init_weights = _init_weights
unslothai/unsloth-zoo/unsloth_zoo/empty_model.pySiglipVisionModel._original_initialize_weights = SiglipVisionModel._init_weights
unslothai/unsloth-zoo/unsloth_zoo/gradient_checkpointing.pytransformers.modeling_utils._unsloth_old_checkpoint = current
unslothai/unsloth-zoo/unsloth_zoo/gradient_checkpointing.pytransformers.modeling_utils.checkpoint = (
unslothai/unsloth-zoo/unsloth_zoo/gradient_checkpointing.pytransformers.modeling_utils.checkpoint = getattr(
unslothai/unsloth-zoo/unsloth_zoo/gradient_checkpointing.pytransformers.modeling_utils.checkpoint = shim
unslothai/unsloth-zoo/unsloth_zoo/gradient_checkpointing.pytransformers.modeling_utils.checkpoint = unsloth_checkpoint
unslothai/unsloth-zoo/unsloth_zoo/loss_utils.pyUnslothForCausalLMLoss._unsloth_fused_loss = UnslothForCausalLMLoss
unslothai/unsloth-zoo/unsloth_zoo/loss_utils.pytransformers.modeling_utils.PreTrainedModel.loss_function = \
unslothai/unsloth-zoo/unsloth_zoo/mlx/loader.pymodeling_llama.LlamaFlashAttention2 = modeling_llama.LlamaAttention
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/common.pysetattr(modeling_module, name, value)
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/compiled_model_identity.pyPreTrainedModel.is_custom_code = method
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/deepseek_v3_moe.pyDeepseekV3ForCausalLM.__UNSLOTH_SUPPORTS_RETURN_HIDDEN_STATES__ = True
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/deepseek_v3_moe.pyDeepseekV3ForCausalLM.forward = _patched_causal_lm_forward
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gemma3n.pyGemma3nTextAttention.forward = forward
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gemma4.pyGemma4AudioAttention.forward = forward
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gemma4.pyGemma4TextAttention.forward = _make_gemma4_attention_carrier_forward(Gemma4TextAttention.forward)
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gemma4.pyLoraModel._create_and_replace = create_and_replace
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gpt_oss.pyGptOssForCausalLM._unsloth_grpo_patched = True
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gpt_oss.pyGptOssForCausalLM.forward = _patched_causal_lm_forward
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gpt_oss.pyGptOssModel.forward = forward
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gpt_oss.pyGptOssPreTrainedModel._unsloth_init_weights_fixed = True
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gpt_oss.pytransformers.generation.utils.create_masks_for_generate = wrap(transformers.generation.utils.create_masks_for_generate)
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gpt_oss.pytransformers.masking_utils.__patched_causal_mask__ = True
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gpt_oss.pytransformers.masking_utils._old_create_causal_mask = _torch_compile(transformers.masking_utils.create_causal_mask, fullgraph = False, dynamic = True)
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gpt_oss.pytransformers.masking_utils._old_create_sliding_window_causal_mask = _torch_compile(transformers.masking_utils.create_sliding_window_causal_mask, fullgraph = False, dynamic = True)
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gpt_oss.pytransformers.masking_utils.create_causal_mask = wrap(create_causal_mask)
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gpt_oss.pytransformers.masking_utils.create_masks_for_generate = wrap(transformers.masking_utils.create_masks_for_generate)
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gpt_oss.pytransformers.masking_utils.create_sliding_window_causal_mask = wrap(create_sliding_window_causal_mask)
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gpt_oss.pytransformers.models.gpt_oss.modeling_gpt_oss.GptOssExperts = GptOssExpertsBnb4bit
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gpt_oss.pytransformers.models.gpt_oss.modeling_gpt_oss.GptOssExperts = \
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gpt_oss.pytransformers.models.gpt_oss.modeling_gpt_oss.GptOssExperts.forward = experts_forward
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gpt_oss.pytransformers.models.gpt_oss.modeling_gpt_oss.GptOssTopKRouter = GptOssTopKRouter
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gpt_oss.pytransformers.models.gpt_oss.modeling_gpt_oss.GptOssTopKRouter = \
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gpt_oss.pytransformers.models.gpt_oss.modeling_gpt_oss._original_GptOssExperts = \
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gpt_oss.pytransformers.models.gpt_oss.modeling_gpt_oss._original_GptOssTopKRouter = \
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gpt_oss.pytransformers.models.gpt_oss.modeling_gpt_oss.create_causal_mask = transformers.masking_utils.create_causal_mask
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gpt_oss.pytransformers.models.gpt_oss.modeling_gpt_oss.create_sliding_window_causal_mask = transformers.masking_utils.create_sliding_window_causal_mask
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gpt_oss.pytransformers.quantizers.quantizer_mxfp4.Mxfp4HfQuantizer._lazy_import_kernels = lambda *args, **kwargs: triton_kernels
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gpt_oss.pytransformers.quantizers.quantizer_mxfp4.Mxfp4HfQuantizer.is_trainable = lambda *args, **kwargs: True
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gpt_oss.pytransformers.quantizers.quantizer_mxfp4.is_kernels_available = is_kernels_available
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/gpt_oss.pytransformers.quantizers.quantizer_mxfp4.is_kernels_available = lambda *args, **kwargs: False
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/misc.pyHiggsAudioV2Embeddings.forward = forward
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/misc.pyModernBertModel._update_attention_mask = _update_attention_mask_contiguous
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/misc.pyPreTrainedModel.gradient_checkpointing_enable = gradient_checkpointing_enable
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/misc.pytransformers.models.mllama.modeling_mllama.MllamaVisionEncoderLayer = MllamaVisionEncoderLayer
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/misc.pytransformers.utils.import_utils.is_causal_conv1d_available = lambda: False
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/moe_utils.pyPeftModel.load_adapter = patched
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/moe_utils.pyPeftModel.save_pretrained = patched
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/moe_utils_bnb4bit.pymodeling_utils.get_model_conversion_mapping = patched_get_model_conversion_mapping
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/mxfp4.pyPreTrainedModel.save_pretrained = save_pretrained
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/mxfp4.pymodeling_utils._get_device_map = _get_device_map
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/mxfp4.pytransformers.integrations.mxfp4.convert_moe_packed_tensors_cpu = convert_moe_packed_tensors_cpu
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/qwen3_5_moe.pytransformers.models.qwen3_5_moe.modeling_qwen3_5_moe.Qwen3_5MoeExperts._unsloth_lora_extractor_fn = staticmethod(_qwen3_5_lora_extractor)
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/qwen3_moe.pytransformers.models.qwen3_moe.modeling_qwen3_moe.Qwen3MoeExperts._unsloth_lora_extractor_fn = staticmethod(_qwen3_lora_extractor)
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/qwen3_next_moe.pytransformers.models.qwen3_next.modeling_qwen3_next.Qwen3NextExperts._unsloth_lora_extractor_fn = staticmethod(_qwen3_next_lora_extractor)
unslothai/unsloth-zoo/unsloth_zoo/temporary_patches/qwen3_vl_moe.pytransformers.models.qwen3_vl_moe.modeling_qwen3_vl_moe.Qwen3VLMoeTextExperts._unsloth_lora_extractor_fn = staticmethod(_qwen3_vl_lora_extractor)
unslothai/unsloth-zoo/unsloth_zoo/vllm_utils.pyQwen3MoeForCausalLM.is_3d_moe_weight = True
unslothai/unsloth/studio/backend/core/inference/safetensors_pread.pymodeling_utils.safe_open = _original
unslothai/unsloth/studio/backend/core/inference/safetensors_pread.pymodeling_utils.safe_open = _with_pread(current)
unslothai/unsloth/unsloth/import_fixes.pyPreTrainedModel.disable_input_require_grads = _patched_single_disable_input_require_grads
unslothai/unsloth/unsloth/import_fixes.pyPreTrainedModel.enable_input_require_grads = _patched_enable_input_require_grads
unslothai/unsloth/unsloth/import_fixes.pyPreTrainedModel.enable_input_require_grads = _patched_single_enable_input_require_grads
unslothai/unsloth/unsloth/import_fixes.pyPreTrainedModel.post_init = post_init
unslothai/unsloth/unsloth/import_fixes.pymodeling_utils.revert_weight_conversion = revert_weight_conversion
unslothai/unsloth/unsloth/kernels/rms_layernorm.pytransformers.models.llama.modeling_llama.LlamaRMSNorm = LlamaRMSNorm
unslothai/unsloth/unsloth/kernels/rms_layernorm.pytransformers.models.llama.modeling_llama.LlamaRMSNorm = Unsloth_LlamaRMSNorm
unslothai/unsloth/unsloth/kernels/rms_layernorm.pytransformers.models.mllama.modeling_mllama.MllamaTextRMSNorm = MllamaTextRMSNorm
unslothai/unsloth/unsloth/kernels/rms_layernorm.pytransformers.models.mllama.modeling_mllama.MllamaTextRMSNorm = Unsloth_MllamaTextRMSNorm
unslothai/unsloth/unsloth/models/_utils.pyPreTrainedModel._get_key_renaming_mapping = _get_key_renaming_mapping
unslothai/unsloth/unsloth/models/_utils.pyPreTrainedModel._get_key_renaming_mapping = original
unslothai/unsloth/unsloth/models/_utils.pyPreTrainedModel.gradient_checkpointing_enable = gradient_checkpointing_enable
unslothai/unsloth/unsloth/models/_utils.pymodeling_utils.get_model_conversion_mapping = get_model_conversion_mapping
unslothai/unsloth/unsloth/models/_utils.pytransformers.integrations.bitsandbytes.should_convert_module = patched_should_convert_module
unslothai/unsloth/unsloth/models/_utils.pytransformers.trainer.get_model_param_count = get_model_param_count
unslothai/unsloth/unsloth/models/_utils.pytransformers.trainer_pt_utils.get_model_param_count = get_model_param_count
unslothai/unsloth/unsloth/models/_utils.pytransformers.utils.import_utils.is_flash_attn_2_available = (
unslothai/unsloth/unsloth/models/_utils.pytransformers.utils.import_utils.is_flash_attn_2_available = lambda *args, **kwargs: (
unslothai/unsloth/unsloth/models/_utils.pytransformers.utils.is_flash_attn_2_available = lambda *args, **kwargs: False
unslothai/unsloth/unsloth/models/_utils.pytransformers.utils.is_openai_available = _lazy_openai_import_check(
unslothai/unsloth/unsloth/models/_utils.pytransformers.utils.quantization_config.BitsAndBytesConfig.__init__ = _BitsAndBytesConfig__init__
unslothai/unsloth/unsloth/models/cohere.pyCohereAttention.__init__ = eval(init_name)
unslothai/unsloth/unsloth/models/cohere.pyCohereAttention.forward = CohereAttention_fast_forward
unslothai/unsloth/unsloth/models/cohere.pyCohereDecoderLayer.forward = CohereDecoderLayer_fast_forward
unslothai/unsloth/unsloth/models/cohere.pyCohereFlashAttention2.forward = CohereAttention_fast_forward
unslothai/unsloth/unsloth/models/cohere.pyCohereForCausalLM.forward = CausalLM_fast_forward(CohereModel_fast_forward_inference)
unslothai/unsloth/unsloth/models/cohere.pyCohereModel.forward = LlamaModel_fast_forward
unslothai/unsloth/unsloth/models/cohere.pyCohereSdpaAttention.forward = CohereAttention_fast_forward
unslothai/unsloth/unsloth/models/cohere.pyPeftModelForCausalLM.forward = PeftModel_fast_forward
unslothai/unsloth/unsloth/models/cohere.pytransformers.models.cohere.modeling_cohere.CohereRotaryEmbedding = LlamaRotaryEmbedding
unslothai/unsloth/unsloth/models/compressed_tensors_int4.pyPeftModel.save_pretrained = save_pretrained
unslothai/unsloth/unsloth/models/falcon_h1.pyFalconH1Attention.__init__ = eval(init_name)
unslothai/unsloth/unsloth/models/falcon_h1.pyFalconH1Attention.forward = FalconH1Attention_fast_forward
unslothai/unsloth/unsloth/models/falcon_h1.pyFalconH1DecoderLayer.forward = FalconH1DecoderLayer_fast_forward
unslothai/unsloth/unsloth/models/falcon_h1.pyFalconH1ForCausalLM.forward = CausalLM_fast_forward(
unslothai/unsloth/unsloth/models/falcon_h1.pyFalconH1Model.forward = LlamaModel_fast_forward
unslothai/unsloth/unsloth/models/falcon_h1.pyPeftModelForCausalLM.forward = PeftModel_fast_forward
unslothai/unsloth/unsloth/models/falcon_h1.pytransformers.models.falcon_h1.modeling_falcon_h1.FalconH1RMSNorm = Unsloth_FalconH1RMSNorm
unslothai/unsloth/unsloth/models/falcon_h1.pytransformers.models.falcon_h1.modeling_falcon_h1.FalconH1RotaryEmbedding = (
unslothai/unsloth/unsloth/models/gemma.pyGemmaAttention.__init__ = eval(init_name)
unslothai/unsloth/unsloth/models/gemma.pyGemmaAttention.forward = LlamaAttention_fast_forward
unslothai/unsloth/unsloth/models/gemma.pyGemmaDecoderLayer.forward = GemmaDecoderLayer_fast_forward
unslothai/unsloth/unsloth/models/gemma.pyGemmaFlashAttention2.forward = LlamaAttention_fast_forward
unslothai/unsloth/unsloth/models/gemma.pyGemmaForCausalLM.forward = CausalLM_fast_forward(GemmaModel_fast_forward_inference)
unslothai/unsloth/unsloth/models/gemma.pyGemmaModel.forward = LlamaModel_fast_forward
unslothai/unsloth/unsloth/models/gemma.pyGemmaSdpaAttention.forward = LlamaAttention_fast_forward
unslothai/unsloth/unsloth/models/gemma.pyPeftModelForCausalLM.forward = PeftModel_fast_forward
unslothai/unsloth/unsloth/models/gemma.pytransformers.models.gemma.modeling_gemma.GemmaRotaryEmbedding = GemmaFixedRotaryEmbedding
unslothai/unsloth/unsloth/models/gemma2.pyGemma2Attention.__init__ = eval(init_name)
unslothai/unsloth/unsloth/models/gemma2.pyGemma2Attention.forward = Gemma2Attention_fast_forward
unslothai/unsloth/unsloth/models/gemma2.pyGemma2DecoderLayer.forward = Gemma2DecoderLayer_fast_forward
unslothai/unsloth/unsloth/models/gemma2.pyGemma2FlashAttention2.forward = Gemma2Attention_fast_forward
unslothai/unsloth/unsloth/models/gemma2.pyGemma2ForCausalLM.forward = CausalLM_fast_forward(Gemma2Model_fast_forward_inference)
unslothai/unsloth/unsloth/models/gemma2.pyGemma2Model.forward = LlamaModel_fast_forward
unslothai/unsloth/unsloth/models/gemma2.pyGemma2SdpaAttention.forward = Gemma2Attention_fast_forward
unslothai/unsloth/unsloth/models/gemma2.pyPeftModelForCausalLM.forward = PeftModel_fast_forward
unslothai/unsloth/unsloth/models/gemma2.pytransformers.models.gemma2.modeling_gemma2.Gemma2RotaryEmbedding = GemmaFixedRotaryEmbedding
unslothai/unsloth/unsloth/models/granite.pyGraniteAttention.__init__ = eval(init_name)
unslothai/unsloth/unsloth/models/granite.pyGraniteAttention.forward = GraniteAttention_fast_forward
unslothai/unsloth/unsloth/models/granite.pyGraniteDecoderLayer.forward = GraniteDecoderLayer_fast_forward
unslothai/unsloth/unsloth/models/granite.pyGraniteFlashAttention2.forward = GraniteAttention_fast_forward
unslothai/unsloth/unsloth/models/granite.pyGraniteForCausalLM.__init__ = patched_init(GraniteForCausalLM.__init__)
unslothai/unsloth/unsloth/models/granite.pyGraniteForCausalLM.forward = CausalLM_fast_forward(GraniteModel_fast_forward_inference)
unslothai/unsloth/unsloth/models/granite.pyGraniteModel.forward = LlamaModel_fast_forward
unslothai/unsloth/unsloth/models/granite.pyGraniteSdpaAttention.forward = GraniteAttention_fast_forward
unslothai/unsloth/unsloth/models/granite.pyPeftModelForCausalLM.forward = PeftModel_fast_forward
unslothai/unsloth/unsloth/models/granite.pytransformers.models.granite.modeling_granite.GraniteRotaryEmbedding = GraniteRotaryEmbedding
unslothai/unsloth/unsloth/models/llama.pyFastLlamaModel.from_pretrained = staticmethod(
unslothai/unsloth/unsloth/models/llama.pyFastLlamaModel.get_peft_model = staticmethod(
unslothai/unsloth/unsloth/models/llama.pyLlamaAttention.__init__ = eval(init_name)
unslothai/unsloth/unsloth/models/llama.pyLlamaAttention.forward = LlamaAttention_fast_forward
unslothai/unsloth/unsloth/models/llama.pyLlamaDecoderLayer.forward = LlamaDecoderLayer_fast_forward
unslothai/unsloth/unsloth/models/llama.pyLlamaFlashAttention2.forward = LlamaAttention_fast_forward
unslothai/unsloth/unsloth/models/llama.pyLlamaForCausalLM.forward = CausalLM_fast_forward(LlamaModel_fast_forward_inference)
unslothai/unsloth/unsloth/models/llama.pyLlamaModel.forward = LlamaModel_fast_forward
unslothai/unsloth/unsloth/models/llama.pyLlamaModel_fast_forward_inference_custom._unsloth_graph_attention = (
unslothai/unsloth/unsloth/models/llama.pyLlamaSdpaAttention.forward = LlamaAttention_fast_forward
unslothai/unsloth/unsloth/models/llama.pyPeftModelForCausalLM.forward = PeftModel_fast_forward
unslothai/unsloth/unsloth/models/llama.pytransformers.models.llama.modeling_llama.LlamaLinearScalingRotaryEmbedding = (
unslothai/unsloth/unsloth/models/llama.pytransformers.models.llama.modeling_llama.LlamaRotaryEmbedding = LlamaRotaryEmbedding
unslothai/unsloth/unsloth/models/loader.pyPreTrainedModel.from_pretrained = classmethod(wrapped)
unslothai/unsloth/unsloth/models/mistral.pyMistralAttention.__init__ = eval(init_name)
unslothai/unsloth/unsloth/models/mistral.pyMistralAttention.forward = MistralAttention_fast_forward
unslothai/unsloth/unsloth/models/mistral.pyMistralDecoderLayer.forward = LlamaDecoderLayer_fast_forward
unslothai/unsloth/unsloth/models/mistral.pyMistralFlashAttention2.forward = MistralAttention_fast_forward
unslothai/unsloth/unsloth/models/mistral.pyMistralForCausalLM.forward = MistralForCausalLM_fast_forward
unslothai/unsloth/unsloth/models/mistral.pyMistralModel.forward = LlamaModel_fast_forward
unslothai/unsloth/unsloth/models/mistral.pyMistralSdpaAttention.forward = MistralAttention_fast_forward
unslothai/unsloth/unsloth/models/mistral.pyPeftModelForCausalLM.forward = PeftModel_fast_forward
unslothai/unsloth/unsloth/models/mistral.pytransformers.models.mistral.modeling_mistral.MistralRotaryEmbedding = LlamaRotaryEmbedding
unslothai/unsloth/unsloth/models/qwen2.pyPeftModelForCausalLM.forward = PeftModel_fast_forward
unslothai/unsloth/unsloth/models/qwen2.pyQwen2Attention.__init__ = eval(init_name)
unslothai/unsloth/unsloth/models/qwen2.pyQwen2Attention.forward = LlamaAttention_fast_forward
unslothai/unsloth/unsloth/models/qwen2.pyQwen2DecoderLayer.forward = LlamaDecoderLayer_fast_forward
unslothai/unsloth/unsloth/models/qwen2.pyQwen2FlashAttention2.forward = LlamaAttention_fast_forward
unslothai/unsloth/unsloth/models/qwen2.pyQwen2ForCausalLM.forward = CausalLM_fast_forward(LlamaModel_fast_forward_inference)
unslothai/unsloth/unsloth/models/qwen2.pyQwen2Model.forward = LlamaModel_fast_forward
unslothai/unsloth/unsloth/models/qwen2.pyQwen2SdpaAttention.forward = LlamaAttention_fast_forward
unslothai/unsloth/unsloth/models/qwen2.pytransformers.models.qwen2.modeling_qwen2.Qwen2RotaryEmbedding = LlamaRotaryEmbedding
unslothai/unsloth/unsloth/models/qwen3.pyPeftModelForCausalLM.forward = PeftModel_fast_forward
unslothai/unsloth/unsloth/models/qwen3.pyQwen3Attention.__init__ = eval(init_name)
unslothai/unsloth/unsloth/models/qwen3.pyQwen3Attention.forward = Qwen3Attention_fast_forward
unslothai/unsloth/unsloth/models/qwen3.pyQwen3DecoderLayer.forward = LlamaDecoderLayer_fast_forward
unslothai/unsloth/unsloth/models/qwen3.pyQwen3FlashAttention2.forward = Qwen3Attention_fast_forward
unslothai/unsloth/unsloth/models/qwen3.pyQwen3ForCausalLM.forward = CausalLM_fast_forward(
unslothai/unsloth/unsloth/models/qwen3.pyQwen3Model.forward = LlamaModel_fast_forward
unslothai/unsloth/unsloth/models/qwen3.pyQwen3SdpaAttention.forward = Qwen3Attention_fast_forward
unslothai/unsloth/unsloth/models/qwen3.pytransformers.models.qwen3.modeling_qwen3.Qwen3RotaryEmbedding = LlamaRotaryEmbedding
unslothai/unsloth/unsloth/models/qwen3_moe.pyPeftModelForCausalLM.forward = PeftModel_fast_forward
unslothai/unsloth/unsloth/models/qwen3_moe.pyQwen3MoeAttention.__init__ = eval(init_name)
unslothai/unsloth/unsloth/models/qwen3_moe.pyQwen3MoeAttention.forward = Qwen3Attention_fast_forward
unslothai/unsloth/unsloth/models/qwen3_moe.pyQwen3MoeDecoderLayer.forward = Qwen3MoeDecoderLayer_fast_forward
unslothai/unsloth/unsloth/models/qwen3_moe.pyQwen3MoeForCausalLM.forward = CausalLM_fast_forward(LlamaModel_fast_forward_inference)
unslothai/unsloth/unsloth/models/qwen3_moe.pyQwen3MoeMLP.forward = fast_swiglu_inference # This is analogous to Dense models' MLP
unslothai/unsloth/unsloth/models/qwen3_moe.pyQwen3MoeModel.forward = LlamaModel_fast_forward
unslothai/unsloth/unsloth/models/qwen3_moe.pytransformers.models.qwen3_moe.modeling_qwen3_moe.Qwen3MoeRotaryEmbedding = (
unslothai/unsloth/unsloth/models/sentence_transformer.pyAutoModel.from_pretrained = original_model_from_pretrained
unslothai/unsloth/unsloth/models/sentence_transformer.pyAutoModel.from_pretrained = return_existing_model
unslothai/unsloth/unsloth/models/sentence_transformer.pymodeling_distilbert.DistilBertModel.forward = forward
unslothai/unsloth/unsloth/models/sentence_transformer.pymodeling_mpnet.MPNetEncoder.forward = forward
unslothai/unsloth/unsloth/models/sentence_transformer.pymodeling_mpnet.MPNetModel._set_gradient_checkpointing = _set_gradient_checkpointing
unslothai/unsloth/unsloth/models/sentence_transformer.pymodeling_mpnet.MPNetModel.supports_gradient_checkpointing = True
unslothai/unsloth/unsloth/models/sentence_transformer.pytransformers.AutoModelForCausalLM = original_causal
unslothai/unsloth/unsloth/models/sentence_transformer.pytransformers.AutoModelForCausalLM = transformers.AutoModel
unslothai/unsloth/unsloth/models/tp_native.pyPreTrainedModel.from_pretrained = classmethod(from_pretrained)
unslothai/unsloth/unsloth/models/tp_native.pyPreTrainedModel.from_pretrained = descriptor
unslothai/unsloth/unsloth/models/vision.pyFastBaseModel.from_pretrained = staticmethod(
unslothai/unsloth/unsloth/models/vision.pyFastBaseModel.get_peft_model = staticmethod(
unslothai/unsloth/unsloth/models/vision.pyPreTrainedModel.to = original_to
unslothai/unsloth/unsloth/models/vision.pyPreTrainedModel.to = to
verl-project/verl/verl/models/mcore/patch.pyGPTModel._postprocess = _patched
verl-project/verl/verl/models/mcore/patch.pyMLASelfAttention.get_query_key_value_tensors = patch_get_query_key_value_tensors
verl-project/verl/verl/models/mcore/patch.pyMultiLatentAttention.forward = patch_forward
verl-project/verl/verl/models/transformers/monkey_patch.pyAutoModelForCausalLMWithValueHead.state_dict = state_dict
verl-project/verl/verl/models/transformers/monkey_patch.pyGlm4vModel.forward = glm4v_base_forward
verl-project/verl/verl/models/transformers/monkey_patch.pyGlm4vTextAttention.forward = glm4v_attn_forward
verl-project/verl/verl/models/transformers/monkey_patch.pyQwen2VLAttention.forward = qwen2_vl_attn_forward
verl-project/verl/verl/models/transformers/monkey_patch.pyQwen2VLModel.forward = qwen2_vl_base_forward
verl-project/verl/verl/models/transformers/monkey_patch.pyQwen2_5_VLAttention.forward = qwen2_vl_attn_forward
verl-project/verl/verl/models/transformers/monkey_patch.pyQwen2_5_VLModel.forward = qwen2_vl_base_forward
verl-project/verl/verl/models/transformers/monkey_patch.pyQwen3VLModel.forward = qwen3_vl_base_forward
verl-project/verl/verl/models/transformers/monkey_patch.pyQwen3VLMoeModel.forward = qwen3_vl_base_forward
verl-project/verl/verl/models/transformers/monkey_patch.pyQwen3VLMoeVisionModel.fast_pos_embed_interpolate = fast_pos_embed_interpolate
verl-project/verl/verl/models/transformers/monkey_patch.pyQwen3VLVisionModel.fast_pos_embed_interpolate = fast_pos_embed_interpolate
verl-project/verl/verl/models/transformers/monkey_patch.pyQwen3_5DecoderLayer.forward = qwen3_5_decoder_layer_forward
verl-project/verl/verl/models/transformers/monkey_patch.pyQwen3_5Model.forward = qwen3_5_base_forward
verl-project/verl/verl/models/transformers/monkey_patch.pyQwen3_5MoeDecoderLayer.forward = qwen3_5_decoder_layer_forward
verl-project/verl/verl/models/transformers/monkey_patch.pyQwen3_5MoeModel.forward = qwen3_5_base_forward
verl-project/verl/verl/models/transformers/monkey_patch.pyQwen3_5MoeVisionModel.fast_pos_embed_interpolate = fast_pos_embed_interpolate
verl-project/verl/verl/models/transformers/monkey_patch.pyQwen3_5VisionModel.fast_pos_embed_interpolate = fast_pos_embed_interpolate
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen2.Qwen2MLP.forward = silu_forward_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen2.Qwen2RMSNorm.forward = rms_norm_forward_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen2.apply_rotary_pos_emb = apply_rotary_pos_emb_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen2_5_vl.Qwen2RMSNorm.forward = rms_norm_forward_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen2_5_vl.Qwen2_5_VLMLP.forward = silu_forward_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen2_5_vl.Qwen2_5_VLRMSNorm.forward = rms_norm_forward_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen3.Qwen3MLP.forward = silu_forward_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen3.Qwen3RMSNorm.forward = rms_norm_forward_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen3.apply_rotary_pos_emb = apply_rotary_pos_emb_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen3_5.Qwen3_5RMSNorm.forward = qwen3_next_rms_norm_forward_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen3_5.Qwen3_5RMSNormGated.forward = qwen3_next_rms_norm_forward_gated_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen3_5.apply_rotary_pos_emb = qwen3_next_apply_rotary_pos_emb_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen3_5_moe.Qwen3_5MoeExperts.forward = qwen3_5_moe_experts_forward_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen3_5_moe.Qwen3_5MoeRMSNorm.forward = qwen3_next_rms_norm_forward_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen3_5_moe.Qwen3_5MoeRMSNormGated.forward = qwen3_next_rms_norm_forward_gated_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen3_5_moe.apply_rotary_pos_emb = qwen3_next_apply_rotary_pos_emb_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen3_moe.Qwen3MoeRMSNorm.forward = rms_norm_forward_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen3_moe.Qwen3MoeSparseMoeBlock.forward = qwen3_moe_sparse_moe_block_forward_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen3_moe.apply_rotary_pos_emb = apply_rotary_pos_emb_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen3_next.Qwen3NextRMSNorm.forward = qwen3_next_rms_norm_forward_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen3_next.Qwen3NextRMSNormGated.forward = qwen3_next_rms_norm_forward_gated_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen3_next.Qwen3NextSparseMoeBlock.forward = qwen3_next_sparse_moe_block_forward_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen3_next.apply_rotary_pos_emb = qwen3_next_apply_rotary_pos_emb_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen3_vl.Qwen3VLTextMLP.forward = silu_forward_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen3_vl.Qwen3VLTextRMSNorm.forward = rms_norm_forward_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen3_vl_moe.Qwen3VLMoeTextRMSNorm.forward = rms_norm_forward_npu
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen3_vl_moe.Qwen3VLMoeTextSparseMoeBlock = NPUQwen3VLMoeTextSparseMoeBlock
verl-project/verl/verl/models/transformers/npu_patch.pymodeling_qwen3_vl_moe.apply_rotary_pos_emb = apply_rotary_pos_emb_npu
vllm-project/llm-compressor/examples/quantizing_moe/deepseek_v4_example.pyDeepseekV4PreTrainedModel._keep_in_fp32_modules_strict = set()
vllm-project/llm-compressor/examples/quantizing_moe/deepseek_v4_mixed_mxfp4_mxfp8.pyDeepseekV4PreTrainedModel._keep_in_fp32_modules_strict = set()
vllm-project/llm-compressor/examples/quantizing_moe/deepseek_v4_pro_example.pyDeepseekV4PreTrainedModel._keep_in_fp32_modules_strict = set()
vllm-project/llm-compressor/src/llmcompressor/modeling/moe/linearize.pytransformers.integrations.finegrained_fp8.replace_with_fp8_linear = (
vllm-project/vllm/vllm/model_executor/models/adapters.pyModelForEmbedding.__name__ = _get_pooling_model_name(cls.__name__, "ForEmbedding")
vllm-project/vllm/vllm/model_executor/models/expert_substitution.pyExpertSubstitutionModel.__name__ = model_cls.__name__
vllm-project/vllm/vllm/model_executor/models/expert_substitution.pyExpertSubstitutionModel.__qualname__ = model_cls.__qualname__
vllm-project/vllm/vllm/v1/attention/backends/composite.pyCompositeAttentionBackend.__module__ = module
vllm-project/vllm/vllm/v1/attention/backends/composite.pyCompositeAttentionBackend.__name__ = name
vllm-project/vllm/vllm/v1/attention/backends/composite.pyCompositeAttentionBackend.__qualname__ = name
\ No newline at end of file