repo stringclasses 147 values | number int64 1 172k | title stringlengths 2 476 | body stringlengths 0 5k | url stringlengths 39 70 | state stringclasses 2 values | labels listlengths 0 9 | created_at timestamp[ns, tz=UTC]date 2017-01-18 18:50:08 2026-01-06 07:33:18 | updated_at timestamp[ns, tz=UTC]date 2017-01-18 19:20:07 2026-01-06 08:03:39 | comments int64 0 58 ⌀ | user stringlengths 2 28 |
|---|---|---|---|---|---|---|---|---|---|---|
vllm-project/vllm | 31,567 | [RFC]: Why custom_mask is not exposed on FlashInfer to get more flexible use case? | ### Motivation.
Like what tensorrt-llm does https://github.com/NVIDIA/TensorRT-LLM/blob/6c1abf2d45c77d04121ebe10f6b29abf89373c60/tensorrt_llm/_torch/attention_backend/flashinfer.py#L411C17-L411C28
### Proposed Change.
expose the custom_weight to support use case like relative attention bias
### Feedback Period.
_No response_
### CC List.
_No response_
### Any Other Things.
_No response_
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/31567 | open | [
"RFC"
] | 2025-12-31T06:00:07Z | 2025-12-31T06:00:07Z | 0 | npuichigo |
vllm-project/vllm | 31,564 | [Bug]: Qwen3-VL-8B-Instruct has accuracy issue - Multi modal accuracy issue | ### Your current environment
**Current input format:**
messages = [
{"role": "system", "content": system_prompt},
{
"role": "user",
"content": [
{"type": "text", "text": user_prompt},
{
"type": "image_url",
"image_url": {"url": image_data_uri}
}
]
}
]
**Command:**
python3 -m vllm serve Qwen/Qwen3-VL-8B-Instruct --max-model-len 22528 --gpu-memory-utilization 0.75 --dtype float16 --port 7001 --trust-remote-code --limit-mm-per-prompt.video 0 --mm-encoder-tp-mode data --mm-processor-cache-gb 0 --tensor-parallel-size 1
**Issue:**
I have a ID number in a fax form like 12347777568 and the model has extracted like 1234777568. The model has skipped 7, but we have four 7 are there and the model returns three 7 as output.
**How to fix this?**
1. Can I increase the max pixels like 2048 or something else.
2. Can I tweak the sampling parameter to allowing the repeated tokens (topp-1 and topk - 0.001) like that.
**Current Sampling:**
"top_k": 20,
"top_p": 0.8,
"repetition_penalty": 1.0,
"temperature": 0.0
### 🐛 Describe the bug
How I need to fix this issue?
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/31564 | open | [
"bug"
] | 2025-12-31T05:13:32Z | 2026-01-02T04:29:14Z | 3 | Dineshkumar-Anandan-ZS0367 |
huggingface/lerobot | 2,737 | SARM WITH PI05: Why trainning loss getting more noise? | ### Ticket Type
❓ Technical Question
### Environment & System Info
```Shell
```
### Description
[SARM with pi05 training for folding towel task _ fold_towel_v3_0 – Weights & Biases.pdf](https://github.com/user-attachments/files/24389716/SARM.with.pi05.training.for.folding.towel.task._.fold_towel_v3_0.Weights.Biases.pdf)
### Context & Reproduction
_No response_
### Relevant logs or stack trace
```Shell
```
### Checklist
- [ ] I have searched existing tickets to ensure this isn't a duplicate.
- [ ] I am using the latest version of the `main` branch.
- [ ] I have verified this is not an environment-specific problem.
### Additional Info / Workarounds
_No response_ | https://github.com/huggingface/lerobot/issues/2737 | closed | [
"question",
"training"
] | 2025-12-31T03:20:16Z | 2026-01-02T08:01:25Z | null | xianglunkai |
huggingface/lerobot | 2,736 | Questions about VLA multi-task training. | ### Ticket Type
💡 Feature Request / Improvement
### Environment & System Info
```Shell
- LeRobot version: 0.4.2
- Platform: Linux-5.15.0-52-generic-x86_64-with-glibc2.31
- Python version: 3.10.18
- Huggingface Hub version: 0.35.3
- Datasets version: 4.1.1
- Numpy version: 2.2.6
- FFmpeg version: 6.1.1
- PyTorch version: 2.7.1+cu126
- Is PyTorch built with CUDA support?: True
- Cuda version: 12.6
- GPU model: NVIDIA GeForce RTX 4060 Ti
- Using GPU in script?: <fill in>
- lerobot scripts: ['lerobot-calibrate', 'lerobot-dataset-viz', 'lerobot-edit-dataset', 'lerobot-eval', 'lerobot-find-cameras', 'lerobot-find-joint-limits', 'lerobot-find-port', 'lerobot-imgtransform-viz', 'lerobot-info', 'lerobot-record', 'lerobot-replay', 'lerobot-setup-motors', 'lerobot-teleoperate', 'lerobot-train']
```
### Description
The generalization capability of VLA mainly comes from pre-training based on large-scale data, but fine-tuning with multi-task co-training also yields good results. This point has been discussed in both the SmolVLA paper and on [Discord](https://discord.com/channels/1216765309076115607/1407325244980727850/1422249462025289809).
<img width="1512" height="1058" alt="Image" src="https://github.com/user-attachments/assets/23cdaa22-a605-474a-9460-1c19e6f29e2d" />
However, the current fine-tuning commands and scripts are based on single-task scenarios. I would like to know how to implement multi-task fine-tuning within the lerobot framework. For example, using it on SmolVLA and pi0.5.
### Context & Reproduction
_No response_
### Relevant logs or stack trace
```Shell
```
### Checklist
- [x] I have searched existing tickets to ensure this isn't a duplicate.
- [x] I am using the latest version of the `main` branch.
- [x] I have verified this is not an environment-specific problem.
### Additional Info / Workarounds
_No response_ | https://github.com/huggingface/lerobot/issues/2736 | open | [
"enhancement",
"question",
"examples",
"training"
] | 2025-12-31T03:12:02Z | 2026-01-04T20:02:02Z | null | yquanli |
vllm-project/vllm | 31,555 | [Docs] Feedback for `/en/stable/`MONSTERDOG | ### 📚 The doc issue
[Projets (1).csv](https://github.com/user-attachments/files/24389184/Projets.1.csv)
[Projets.csv](https://github.com/user-attachments/files/24389185/Projets.csv)
[MonsterDog_Pilot_ROI_ISO42001_Report.pdf](https://github.com/user-attachments/files/24389187/MonsterDog_Pilot_ROI_ISO42001_Report.pdf)
[MonsterDog_Pilot_ROI_ISO42001_Report.pdf](https://github.com/user-attachments/files/24389186/MonsterDog_Pilot_ROI_ISO42001_Report.pdf)
[LIVRE_BLANC_MONSTERDOG_VINF.md](https://github.com/user-attachments/files/24389188/LIVRE_BLANC_MONSTERDOG_VINF.md)
[MONSTERDOG_TOTALITY_SUPREME_INFINITY.py](https://github.com/user-attachments/files/24389189/MONSTERDOG_TOTALITY_SUPREME_INFINITY.py)
[SCRIPT_ULTIME_FINAL_vULT_FULL.md](https://github.com/user-attachments/files/24389190/SCRIPT_ULTIME_FINAL_vULT_FULL.md)
[RAPPORT_FINAL_MONSTERDOG.md](https://github.com/user-attachments/files/24389191/RAPPORT_FINAL_MONSTERDOG.md)
<img width="1024" height="1024" alt="Image" src="https://github.com/user-attachments/assets/d7e55288-c704-4bf3-86a1-0d04a8a081a0" />
[safe_hold_v1_1.py](https://github.com/user-attachments/files/24389193/safe_hold_v1_1.py)
[safe_hold_v1_1.py](https://github.com/user-attachments/files/24389192/safe_hold_v1_1.py)
[★MONSTERDOG★OMNI★AEGIS★2026.py](https://github.com/user-attachments/files/24389194/MONSTERDOG.OMNI.AEGIS.2026.py)
### Suggest a potential alternative/fix
[MonsterDog_Pilot_ROI_ISO42001_Report.pdf](https://github.com/user-attachments/files/24389173/MonsterDog_Pilot_ROI_ISO42001_Report.pdf)
[MonsterDog_Pilot_ROI_ISO42001_Report.pdf](https://github.com/user-attachments/files/24389172/MonsterDog_Pilot_ROI_ISO42001_Report.pdf)
[LIVRE_BLANC_MONSTERDOG_VINF.md](https://github.com/user-attachments/files/24389174/LIVRE_BLANC_MONSTERDOG_VINF.md)
[MONSTERDOG_TOTALITY_SUPREME_INFINITY.py](https://github.com/user-attachments/files/24389175/MONSTERDOG_TOTALITY_SUPREME_INFINITY.py)
[SCRIPT_ULTIME_FINAL_vULT_FULL.md](https://github.com/user-attachments/files/24389176/SCRIPT_ULTIME_FINAL_vULT_FULL.md)
[RAPPORT_FINAL_MONSTERDOG.md](https://github.com/user-attachments/files/24389177/RAPPORT_FINAL_MONSTERDOG.md)
[safe_hold_v1_1.py](https://github.com/user-attachments/files/24389178/safe_hold_v1_1.py)
[safe_hold_v1_1.py](https://github.com/user-attachments/files/24389179/safe_hold_v1_1.py)
[★MONSTERDOG★OMNI★AEGIS★2026.py](https://github.com/user-attachments/files/24389180/MONSTERDOG.OMNI.AEGIS.2026.py)
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/31555 | closed | [
"documentation"
] | 2025-12-31T01:20:55Z | 2025-12-31T05:18:48Z | 0 | s33765387-cpu |
huggingface/lerobot | 2,735 | Buy the camera? | Hi! Where do I buy the camera and the whole SO-ARM101 kit?
I find the kit at a chinese website like WoWRobo Robotics with only Paypal payment. But is that it? How do I buy the camera otherwise? | https://github.com/huggingface/lerobot/issues/2735 | open | [
"question",
"sensors"
] | 2025-12-30T22:32:42Z | 2025-12-30T22:51:39Z | null | JFI12 |
huggingface/candle | 3,272 | Added support for Vulkan, any interest? | I have a Intel Arc A770 16GB GPU and wanted to use it with candle.
I took niklasha's work on niklas-vulkan-2 branch cherry-pick's into the current main branch.
I (when I say I, I mean I was the navigator, Codex 5.2 max did the work) added the following:
Added Vulkan queue-family selection and synchronize() so VulkanDevice uses compute-capable queues and can block on GPU work (device.rs).
Expanded Vulkan storage surface with raw_buffer() access for kernel dispatch and fixed error wiring (storage.rs).
Wired Vulkan kernel registry to include matmul, norms, softmax, masked softmax, and quantized kernels (lib.rs).
Added F32/F16 matmul shader stubs and norm/softmax shaders for initial Vulkan ops coverage (*.comp).
Implemented Vulkan masked softmax and staged SDPA path with GQA support in candle-nn (ops.rs).
Added Vulkan smoke tests and masked softmax correctness test (vulkan_smoke_tests.rs, vulkan_masked_softmax.rs).
Fixed missing imports and push-constant binding for Vulkan command execution (storage.rs).
Added bytemuck + vulkano-shaders feature wiring for Vulkan builds (Cargo.toml).
Introduced QVulkanStorage backed by raw byte buffers with dequantize/quantize helpers (vulkan.rs).
Added Vulkan quantized matmul kernels for Q5_0 and Q8_0 (naive, F32 output) (qmatmul_q5_0_f32.comp, qmatmul_q8_0_f32.comp).
Hooked Vulkan quantized path into QTensor forward and added Vulkan quantized tests (mod.rs, vulkan_quantized_tests.rs).
Added a dequantize‑fallback backward path for QLoRA-style gradients (mod.rs).
Cleaned up dummy Vulkan stubs to match new quantized API surface (dummy_vulkan.rs).
Fixed multiple test harness macro/feature mismatches to compile with Vulkan enabled (test_utils.rs, *.rs). | https://github.com/huggingface/candle/issues/3272 | open | [] | 2025-12-30T02:58:27Z | 2025-12-30T03:00:12Z | 0 | davidwynter |
vllm-project/vllm | 31,515 | [Feature]: need scheduler solution with high priority to process prefill | ### 🚀 The feature, motivation and pitch
I have a model situiation which is that the model just care about the throughtput not care about the time delay, so I need a schedule solution which can get the high priority to process prefill and after all prefill is finished in the batch and then process the decode, this solution can increase the decode batch_size at the best. I need this feature to support in vllm ascend~
### Alternatives
_No response_
### Additional context
_No response_
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/31515 | open | [
"feature request"
] | 2025-12-30T02:09:35Z | 2025-12-30T02:09:35Z | 0 | 184603418 |
vllm-project/vllm | 31,486 | [Feature]: GLM 4.7 vocab padding feature | ### 🚀 The feature, motivation and pitch
The number of attention heads in GLM-4.7 is 96, so I’m trying to run the FP8 version with 6× H20 GPUs using tensor parallelism (tp=6).
However, vllm serve fails and due to `151552 cannot be divided by 6`.
This seems to be caused by the vocab size 151552 not being divisible by the TP size. In my understanding, this could be solvable by padding the vocab size up.
Alternatively, is there any simpler workaround or recommended solution for this case? Thanks!
### Alternatives
_No response_
### Additional context
_No response_
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/31486 | open | [
"feature request"
] | 2025-12-29T09:30:35Z | 2026-01-06T02:45:22Z | 3 | H100-H200-B200 |
vllm-project/vllm | 31,484 | [Usage]: RuntimeError when running Qwen2.5-VL-7B-Instruct with vllm: Potential version incompatibility | ### Your current environment
```text
Collecting environment information...
==============================
System Info
==============================
OS : Ubuntu 24.04.2 LTS (x86_64)
GCC version : (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
Clang version : Could not collect
CMake version : Could not collect
Libc version : glibc-2.39
==============================
PyTorch Info
==============================
PyTorch version : 2.9.0+cu128
Is debug build : False
CUDA used to build PyTorch : 12.8
ROCM used to build PyTorch : N/A
==============================
Python Environment
==============================
Python version : 3.12.12 | packaged by Anaconda, Inc. | (main, Oct 21 2025, 20:16:04) [GCC 11.2.0] (64-bit runtime)
Python platform : Linux-6.8.0-53-generic-x86_64-with-glibc2.39
==============================
CUDA / GPU Info
==============================
Is CUDA available : True
CUDA runtime version : 12.8.93
CUDA_MODULE_LOADING set to :
GPU models and configuration :
GPU 0: NVIDIA B200
GPU 1: NVIDIA B200
GPU 2: NVIDIA B200
GPU 3: NVIDIA B200
GPU 4: NVIDIA B200
GPU 5: NVIDIA B200
GPU 6: NVIDIA B200
GPU 7: NVIDIA B200
Nvidia driver version : 570.148.08
cuDNN version : Could not collect
HIP runtime version : N/A
MIOpen runtime version : N/A
Is XNNPACK available : True
==============================
CPU Info
==============================
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 52 bits physical, 57 bits virtual
Byte Order: Little Endian
CPU(s): 144
On-line CPU(s) list: 0-143
Vendor ID: GenuineIntel
BIOS Vendor ID: Intel(R) Corporation
Model name: Intel(R) Xeon(R) 6960P
BIOS Model name: Intel(R) Xeon(R) 6960P CPU @ 2.7GHz
BIOS CPU family: 179
CPU family: 6
Model: 173
Thread(s) per core: 1
Core(s) per socket: 72
Socket(s): 2
Stepping: 1
BogoMIPS: 5400.00
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 monitor ds_cpl smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cat_l2 cdp_l3 intel_ppin cdp_l2 ssbd mba ibrs ibpb stibp ibrs_enhanced fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local split_lock_detect user_shstk avx_vnni avx512_bf16 wbnoinvd dtherm ida arat pln pts hfi avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b enqcmd fsrm md_clear serialize tsxldtrk pconfig arch_lbr ibt amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities
L1d cache: 6.8 MiB (144 instances)
L1i cache: 9 MiB (144 instances)
L2 cache: 288 MiB (144 instances)
L3 cache: 864 MiB (2 instances)
NUMA node(s): 2
NUMA node0 CPU(s): 0-71
NUMA node1 CPU(s): 72-143
Vulnerability Gather data sampling: Not affected
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Not affected
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed: Not affected
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Enhanced / Automatic IBRS; IBPB conditional; RSB filling; PBRSB-eIBRS Not affected; BHI BHI_DIS_S
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
========== | https://github.com/vllm-project/vllm/issues/31484 | open | [
"usage"
] | 2025-12-29T08:36:11Z | 2025-12-30T02:40:38Z | 1 | puyuan1996 |
huggingface/diffusers | 12,899 | Training script of z-image controlnet? | Can diffusers provide training script of z-image controlnet? | https://github.com/huggingface/diffusers/issues/12899 | open | [] | 2025-12-29T08:30:09Z | 2025-12-29T08:30:09Z | 0 | universewill |
vllm-project/vllm | 31,480 | [Usage]: run deepseek v3.2 failed | ### Your current environment
Collecting environment information...
==============================
System Info
==============================
OS : Ubuntu 22.04.5 LTS (x86_64)
GCC version : (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version : Could not collect
CMake version : version 3.22.1
Libc version : glibc-2.35
==============================
PyTorch Info
==============================
PyTorch version : 2.9.0+cu128
Is debug build : False
CUDA used to build PyTorch : 12.8
ROCM used to build PyTorch : N/A
==============================
Python Environment
==============================
Python version : 3.12.12 | packaged by Anaconda, Inc. | (main, Oct 21 2025, 20:16:04) [GCC 11.2.0] (64-bit runtime)
Python platform : Linux-5.15.0-78-generic-x86_64-with-glibc2.35
==============================
CUDA / GPU Info
==============================
Is CUDA available : True
CUDA runtime version : 12.8.93
CUDA_MODULE_LOADING set to :
GPU models and configuration :
GPU 0: NVIDIA RTX PRO 6000 Blackwell Server Edition
GPU 1: NVIDIA RTX PRO 6000 Blackwell Server Edition
GPU 2: NVIDIA RTX PRO 6000 Blackwell Server Edition
GPU 3: NVIDIA RTX PRO 6000 Blackwell Server Edition
GPU 4: NVIDIA RTX PRO 6000 Blackwell Server Edition
GPU 5: NVIDIA RTX PRO 6000 Blackwell Server Edition
GPU 6: NVIDIA RTX PRO 6000 Blackwell Server Edition
GPU 7: NVIDIA RTX PRO 6000 Blackwell Server Edition
Nvidia driver version : 580.95.05
cuDNN version : Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.9.8.0
/usr/lib/x86_64-linux-gnu/libcudnn_adv.so.9.8.0
/usr/lib/x86_64-linux-gnu/libcudnn_cnn.so.9.8.0
/usr/lib/x86_64-linux-gnu/libcudnn_engines_precompiled.so.9.8.0
/usr/lib/x86_64-linux-gnu/libcudnn_engines_runtime_compiled.so.9.8.0
/usr/lib/x86_64-linux-gnu/libcudnn_graph.so.9.8.0
/usr/lib/x86_64-linux-gnu/libcudnn_heuristic.so.9.8.0
/usr/lib/x86_64-linux-gnu/libcudnn_ops.so.9.8.0
HIP runtime version : N/A
MIOpen runtime version : N/A
Is XNNPACK available : True
==============================
CPU Info
==============================
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 52 bits physical, 57 bits virtual
Byte Order: Little Endian
CPU(s): 208
On-line CPU(s) list: 0-207
Vendor ID: GenuineIntel
Model name: Intel(R) Xeon(R) Platinum 8470Q
CPU family: 6
Model: 143
Thread(s) per core: 2
Core(s) per socket: 52
Socket(s): 2
Stepping: 8
CPU max MHz: 3800.0000
CPU min MHz: 800.0000
BogoMIPS: 4200.00
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cat_l2 cdp_l3 invpcid_single intel_ppin cdp_l2 ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local split_lock_detect avx_vnni avx512_bf16 wbnoinvd dtherm ida arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b enqcmd fsrm md_clear serialize tsxldtrk pconfig arch_lbr amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities
Virtualization: VT-x
L1d cache: 4.9 MiB (104 instances)
L1i cache: 3.3 MiB (104 instances)
L2 cache: 208 MiB (104 instances)
L3 cache: 210 MiB (2 instances)
NUMA node(s): 2
NUMA node0 CPU(s): 0-51,104-155
NUMA node1 CPU(s): 52-103,156-207
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Not affected
Vulnerability Retbleed: Not affected
Vulnerabili | https://github.com/vllm-project/vllm/issues/31480 | open | [
"usage"
] | 2025-12-29T07:33:04Z | 2025-12-29T07:33:04Z | 0 | ljwps |
vllm-project/vllm | 31,479 | [Feature]: Enable LoRA support for tower and connector in more MM models | ### 🚀 The feature, motivation and pitch
Regarding multi-modal models, we have supported adding LoRA to the tower encoder and connector,see: #26674, but have only implemented it for a few models (`Qwen VL series` and `idefics3`). There is no reason not to support other multi-modal models.
### Solution
For the remaining models we want to support adding LoRA to the tower encoder and connector, we need to implement the following 2 functions:
`get_num_mm_encoder_tokens`
`get_num_mm_connector_tokens`
**The root cause we need to implement these two functions is:** the number of multi-modal tokens represented in the language model does not necessarily match the input length required by the linear layers in the vision tower or connector. Since the lora_mapping requires the precise input token length prior to activation, these helper functions are necessary to bridge the discrepancy and calculate the correct lengths.
### List of models that are completed or WIP
- Qwen VL series: #26674
- idefics3: #26674
- LLaVA: https://github.com/vllm-project/vllm/pull/31513
- BLIP2: https://github.com/vllm-project/vllm/pull/31620
- GLM4 : https://github.com/vllm-project/vllm/pull/31652
- PaliGemma https://github.com/vllm-project/vllm/pull/31656
- H2OVL https://github.com/vllm-project/vllm/pull/31696
- Pixtral https://github.com/vllm-project/vllm/pull/31724
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/31479 | open | [
"help wanted",
"feature request"
] | 2025-12-29T07:28:52Z | 2026-01-06T02:03:29Z | 4 | jeejeelee |
vllm-project/vllm | 31,474 | [Feature]: GLM 4.7 vocab padding feature | ### 🚀 The feature, motivation and pitch
The number of attention heads in GLM-4.7 is 96, so I’m trying to run the FP8 version with 6× H20 GPUs using tensor parallelism (tp=6).
However, vllm serve fails and due to `151552 cannot be divided by 6`.
This seems to be caused by the vocab size 151552 not being divisible by the TP size. In my understanding, this could be solvable by padding the vocab size up.
Alternatively, is there any simpler workaround or recommended solution for this case? Thanks!
### Alternatives
_No response_
### Additional context
_No response_
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/31474 | closed | [
"feature request"
] | 2025-12-29T04:55:28Z | 2025-12-29T09:28:17Z | 0 | H100-H200-B200 |
vllm-project/vllm | 31,469 | [Feature]: Optimize the definition of the fake function in the code. | ### 🚀 The feature, motivation and pitch
The current code contains some fake function definitions, which are placed together with the main logic, such as `all_reduce_fake`. In the `parallel_state.py` file, can we define a file called `parallel_state_fake.py` and move all the corresponding fake functions to this file, and do the same for the others?
### Alternatives
_No response_
### Additional context
_No response_
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/31469 | open | [
"feature request"
] | 2025-12-29T03:14:26Z | 2025-12-29T06:16:08Z | 3 | lengrongfu |
vllm-project/vllm | 31,467 | [RFC]: A Triton operator dispatch mechanism through modified `CustomOp` | ### Motivation.
Triton is becoming increasingly important in vLLM, and we've noticed its use in many models, quantization processes, and general workflows. Meanwhile, vLLM supports various backends. Typically, to achieve high performance, **different implementations of the Triton kernels** are used on different hardware, such as Ascend NPU. However, we've observed that vLLM currently lacks an effective operator dispatch mechanism for Triton to ensure that various backends can implement their own Triton kernels, which are then uniformly called by vLLM.
There are 3 ways of calling triton function now:
#### Through Attention Backend
Triton functions are called in `Attention` layer when the attention backend is specified as `TRITON_ATTN` or `TRITON_MLA`.
```python
current_platform.get_attn_backend_cls(...)
```
#### Through CustomOp
Some triton functions are included in other customops's forward pipeline, and they are put into `forward_cuda`, e.g., `causal_conv1d_fn` and `causal_conv1d_update` in `ShortConv`.
```python
class op1(CustomOp):
def forward_cuda(kwargs):
triton_fn(**kwargs)
```
#### Directly call
And there are others directly call triton functions in the normal pipeline.
- some models derictly call triton functions in forward
- Qwen3-Next
- Kimi-Linear
- ...
- modelrunner v2
- block table
- input batch
Also, I notice that the implements are different form rocm and nvidia, algouth they are both cuda-alike platform.
```python
if current_platform.is_rocm():
@triton.jit
def round_int8(x):
return tl.extra.hip.libdevice.round(x).to(tl.int8)
else:
@triton.jit
def round_int8(x):
return tl.extra.cuda.libdevice.round(x).to(tl.int8)
```
### Proposed Change.
To solve the issues above, we propose to do the following changes:
<img width="1537" height="1346" alt="Image" src="https://github.com/user-attachments/assets/425dd6a4-b15c-4057-8aa8-25e4f563970b" />
1. Abstract a `CustomOpBase` class, which maintains funtions `register`, `register_oot` and `forward_dispatch`, which means all the instance of `CustomOpBase` could be registered in/out of vllm.
2. Seperate `CustomOp` and `CustomTritonOp`, we dispatch `CustomTritonOp` in a python func level, which pairs with the triton kernel. And the `CustomOp` keeps as is.
3. Refactor the exsiting triton kernels that are directly called without a python funtion warpping it, e.g., `eagle_prepare_inputs_padded_kernel`
4. Refactor the triton python functions to be hierit from `CustomTritonOp`, and optimize the current implement of triton kernel patching.
#### Example
##### Code Change
```python
class CustomOpBase:
"""
Base class for custom op. This class mainly offer the registry and dispatch function,
and others must be overwrite in the sub classes.
Dispatches the forward method to the appropriate backend.
"""
op_registry: dict[str, Any] = {}
op_registry_oot: dict[str, Any] = {}
def __new__(cls, *args, **kwargs):
try:
op_name = cls.__name__
except AttributeError:
raise TypeError(
f"Cannot instantiate '{cls.__name__}': its 'name' attribute "
f"was not set, possibly because it was not decorated with "
f"@CustomOp.register, or it's the CustomOp base class itself."
) from None
if op_name not in cls.op_registry_oot:
op_cls_to_instantiate = cls
else:
op_cls_to_instantiate = cls.op_registry_oot[op_name]
logger.debug(
"Instantiating custom op: %s using %s",
op_name,
str(op_cls_to_instantiate),
)
return super().__new__(op_cls_to_instantiate)
def __init__(self, enforce_enable: bool = False):
self._enforce_enable = enforce_enable
self._forward_method = self.dispatch_forward()
def forward(self, *args, **kwargs):
return self._forward_method(*args, **kwargs)
def forward_native(self, *args, **kwargs):
raise NotImplementedError
def forward_cuda(self, *args, **kwargs):
raise NotImplementedError
def forward_x(self, *args, **kwargs):
raise NotImplementedError
def forward_oot(self, *args, **kwargs):
raise NotImplementedError
def dispatch_forward(self):
raise NotImplementedError
# Decorator to register custom ops.
@classmethod
def register(cls, name: str):
def decorator(op_cls):
assert name not in cls.op_registry, f"Duplicate op name: {name}"
op_cls.name = name
cls.op_registry[name] = op_cls
return op_cls
return decorator
@classmethod
def register_oot(cls, _decorated_op_cls=None, name: str | None = None):
def decorator(op_cls):
reg_name = name if name is not None else cls.__name__
assert reg_name not in cls.op_registry_oot, f"Duplicate op name: {reg_ | https://github.com/vllm-project/vllm/issues/31467 | open | [
"RFC"
] | 2025-12-29T02:44:13Z | 2026-01-06T07:38:29Z | 12 | MengqingCao |
vllm-project/vllm | 31,437 | [Bug]: Streaming tool calls missing id/type/name in finish chunk | ### Your current environment
vLLM 0.14.0rc1.dev3 (but also affects main branch as of today)
### Model
GLM-4.7-AWQ with `--tool-call-parser glm47` (also affects other parsers that emit complete tool calls)
### What is the issue?
When streaming tool calls, the finish chunk code in `serving_chat.py` overwrites the tool parser's properly-formatted `DeltaMessage` with a stripped-down version that only contains `index` and `function.arguments`, losing the `id`, `type`, and `function.name` fields.
This breaks OpenAI-compatible clients that expect `id` to be present in tool call responses.
### Root cause
In `serving_chat.py` around line 1237, when `_should_check_for_unstreamed_tool_arg_tokens()` returns true:
```python
remaining_call = expected_call.replace(actual_call, "", 1)
delta_message = DeltaMessage(
tool_calls=[
DeltaToolCall(
index=index,
function=DeltaFunctionCall(
arguments=remaining_call
).model_dump(exclude_none=True),
)
]
)
```
This creates a new `DeltaMessage` without preserving `id`, `type`, or `function.name` from the original `delta_message` that the tool parser returned.
### Proposed fix
Preserve the fields from the original delta:
```python
remaining_call = expected_call.replace(actual_call, "", 1)
original_tc = delta_message.tool_calls[0]
original_fn = original_tc.function if original_tc else None
delta_message = DeltaMessage(
tool_calls=[
DeltaToolCall(
index=index,
id=original_tc.id if original_tc else None,
type=original_tc.type if original_tc else None,
function=DeltaFunctionCall(
name=original_fn.name if original_fn else None,
arguments=remaining_call,
),
)
]
)
```
### Why this wasn't caught before
This code path only triggers when the tool parser hasn't streamed all argument tokens yet. Many parsers stream arguments incrementally, so they rarely hit this path. Parsers like GLM that emit complete tool calls at once trigger it consistently.
### Related issues
- #16340 (similar symptoms, different root cause)
- #10781 (mentions delta not being submitted correctly)
Happy to submit a PR if this approach looks right.
### Before submitting a new issue...
- [X] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/31437 | closed | [] | 2025-12-27T23:54:20Z | 2025-12-29T13:10:54Z | 0 | amittell |
vllm-project/vllm | 31,414 | [Feature][Cleanup]: Unify `vllm.utils.flashinfer` and `vllm.model_executor.layers.quantization.utils.flashinfer_utils` | ### 🚀 The feature, motivation and pitch
its confusing to have both
### Alternatives
_No response_
### Additional context
_No response_
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/31414 | open | [
"help wanted",
"good first issue",
"feature request"
] | 2025-12-27T18:27:00Z | 2025-12-31T22:25:36Z | 4 | robertgshaw2-redhat |
vllm-project/vllm | 31,398 | [Doc]: Eagle3 with tensor parallelism | ### 📚 The doc issue
According to https://docs.vllm.ai/en/latest/features/spec_decode/#speculating-using-eagle-based-draft-models:
> The EAGLE based draft models need to be run without tensor parallelism (i.e. draft_tensor_parallel_size is set to 1 in speculative_config), although it is possible to run the main model using tensor parallelism (see example above).
But there's no explanation for why the draft tpsize could only be set to 1, so I checked the code and found:
https://github.com/vllm-project/vllm/blob/52bf0665168c539d2d061a664ad62b18a12e80bb/vllm/config/speculative.py#L441-L447
and
https://github.com/vllm-project/vllm/blob/52bf0665168c539d2d061a664ad62b18a12e80bb/vllm/config/speculative.py#L563-L571
I did not find any explicit restriction that enforces the draft model to run without tensor parallelism.
So I guess the `draft_tensor_parallel_size` should be set to **either** 1 **or** the same value as the target_model. And also I tried doing so, and found that the tensor parallelism seems worked correctly.
Is it possible that this functionality has already been implemented, but the documentation has not been updated accordingly?
### Suggest a potential alternative/fix
Just change one line of documentation as mentioned above:
> It's possible to run the EAGLE based draft models with tensor_parallel using tp_size=1 or target_model_tpsize (i.e. `draft_tensor_parallel_size` is set to either 1 or the same value as the target_model in speculative_config).
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/31398 | open | [
"documentation"
] | 2025-12-27T03:10:50Z | 2026-01-04T01:21:07Z | 3 | JSYRD |
huggingface/transformers | 43,048 | Need to understand difference between TP support via transformers code v/s Pytorch's native parallelize_module API. | Based on the existing code base of transformers, below sequence of operations are performed on model object to make it TP compatible.
- TP Plan for Llama: https://github.com/huggingface/transformers/blob/a7f29523361b2cc12e51c1f5133d95f122f6f45c/src/transformers/models/llama/configuration_llama.py#L113
- self._tp_plan populated based on above default plan:
https://github.com/huggingface/transformers/blob/a7f29523361b2cc12e51c1f5133d95f122f6f45c/src/transformers/modeling_utils.py#L1325
- from_pretrained calls distribute_model
https://github.com/huggingface/transformers/blob/a7f29523361b2cc12e51c1f5133d95f122f6f45c/src/transformers/modeling_utils.py#L3944
- distribute_model internally applies TP hooks based on the plans defined for each module.
https://github.com/huggingface/transformers/blob/a7f29523361b2cc12e51c1f5133d95f122f6f45c/src/transformers/integrations/tensor_parallel.py#L1307
I want to understand how this is different than parallelize_module API of Pytorch (https://docs.pytorch.org/docs/stable/distributed.tensor.parallel.html#torch.distributed.tensor.parallel.parallelize_module).
One example of TP+DP can be referred from below link.
https://github.com/pytorch/pytorch/blob/7de041cb5a5817500b973eb32a70325187a83407/test/distributed/_composable/test_composability/test_2d_composability.py#L478
From the Pytorch example, it looks very clean to work with plain DP and TP. But when using Transformer's Trainer along with Accelerate for Plain DP+TP then there are lot of complications identified in https://github.com/huggingface/accelerate/issues/3876#issuecomment-3627324602.
I would like to understand the difference between the existing transformers approach and plain Pytorch approach and request streamlining the implementation of transformers as well as accelerate if that feels suitable. | https://github.com/huggingface/transformers/issues/43048 | open | [] | 2025-12-26T10:05:38Z | 2026-01-05T15:35:13Z | 1 | quic-meetkuma |
huggingface/lerobot | 2,721 | The virtual machine is unable to recognize the keyboard. | ### Ticket Type
❓ Technical Question
### Environment & System Info
```Shell
(base) tom@tom-VMware-Virtual-Platform:~/lerobot_alohamini$ python check_lerobot.py
使用现有的DISPLAY: :0
=== 环境诊断 ===
Python 版本: 3.12.12 | packaged by conda-forge | (main, Oct 22 2025, 23:25:55) [GCC 14.3.0]
DISPLAY 环境变量: :0
XDG_SESSION_TYPE 环境变量: wayland
Wayland_DISPLAY 环境变量: 未设置
===============
正在启动键盘监听器...
请尝试按下一些字母键和方向键。
按 `ESC` 键退出测试。
监听器线程已启动。等待按键输入...
wsdasdwsdasdfdaswdsdfawdsa
```
### Description
When you use the Ubuntu system of the virtual machine to control the main arm and chassis, you may encounter a problem where the keyboard cannot be recognized. This problem is actually quite easy to solve. All you need to do is log out of your desktop, go to the login screen, and click the ⚙ gear icon below the username to select "Ubuntu on Xorg". The reason for this problem is that the pynput library relies on the X11 protocol, while Wayland is a new display server protocol, and the two are not fully compatible.After that, you can safely use your keyboard.
### Context & Reproduction
_No response_
### Relevant logs or stack trace
```Shell
```
### Checklist
- [ ] I have searched existing tickets to ensure this isn't a duplicate.
- [ ] I am using the latest version of the `main` branch.
- [ ] I have verified this is not an environment-specific problem.
### Additional Info / Workarounds
_No response_ | https://github.com/huggingface/lerobot/issues/2721 | open | [
"question"
] | 2025-12-26T08:02:27Z | 2025-12-26T08:02:37Z | null | ht202 |
huggingface/transformers | 43,045 | Multimodal chat sample | ### Feature request
Add a sample covering chat scenario including images, videos or audio.
### Motivation
`AutoModelForCausalLM`'s `use_cache` is barely documented.
Describe a pattern handling the following cases
1. Tokenizer replaces tokens that are already in kv cache with a different token. For example, the model generated 2 tokens with string representations: `a` and `b` and the tokenizer replaces them with a single `a b` token on the next iteration invalidating a part of kv cache
2. Reuse embeddings computed earlier for non text modalities
There's https://github.com/huggingface/transformers/blob/a7f29523361b2cc12e51c1f5133d95f122f6f45c/src/transformers/cli/chat.py but it doesn't cover non text modalities.
### Your contribution
I'm fine to submit a PR. That will help me to learn along the way. But I need guidance how to resolve the issues I described in the motivation section. | https://github.com/huggingface/transformers/issues/43045 | closed | [
"Feature request"
] | 2025-12-26T06:16:53Z | 2025-12-31T10:36:38Z | 9 | Wovchena |
sgl-project/sglang | 15,860 | [Ask for help] How to deploy GLM-4.7 | Hi, can anyone help me to deploy GLM-4.7? I encounter a bug when using `sglang==0.5.6.post2` (which is latest on `https://github.com/sgl-project/sglang`). What is the correct version for GLM-4.7?
```
launch_server.py: error: argument --tool-call-parser: invalid choice: 'glm47' (choose from 'deepseekv3', 'deepseekv31', 'deepseekv32', 'glm', 'glm45', 'gpt-oss', 'kimi_k2', 'llama3', 'mistral', 'pythonic', 'qwen', 'qwen25', 'qwen3_coder', 'step3', 'minimax-m2')
```
Thanks so much!!!!!!!!!!!
<img width="823" height="229" alt="Image" src="https://github.com/user-attachments/assets/241144fa-2038-4f43-be6d-de0895071ffe" /> | https://github.com/sgl-project/sglang/issues/15860 | open | [] | 2025-12-26T02:59:06Z | 2025-12-28T21:21:17Z | 2 | sunjie279 |
huggingface/tokenizers | 1,919 | De/tokenization on CUDA | Could at least de-tokenization be done directly on CUDA? Like in my hack `bpedecode_vec` in https://github.com/pytorch/pytorch/issues/135704#issue-2520180382 which indexes into a detokenization vocab byte table via `repeat_interleave`
Also, maybe for better CUDAGraph-ability / no CPU syncs, there should be some static-sized pre-allocated `out=` version, like `torch.nonzero_static`?
---
Offtopic: it's also a bit inconsistent naming to have `batch_decode` and `batch_encode_plus`... What is the motivation for the `_plus` suffix? | https://github.com/huggingface/tokenizers/issues/1919 | open | [] | 2025-12-26T02:20:49Z | 2026-01-05T10:51:17Z | 1 | vadimkantorov |
vllm-project/vllm | 31,361 | [Usage]: Question about the dummy run。It seems the dummy run use different precision? | ### Question
I am trying to modify vllm. especially the **tp** communication, i'am tring to **break all-reduce into reduce-scatter + all-gather**.
However I encountered precision problem, after i print the hidden states. it seems each layer has around +-0.01 diff, when it accumulated over all the layers, the result seems to be a huge difference. I thought it may be my implementation error. But after I checked the log, I see some dummy run before executing real request. **I checked the dummy run's data. It perfectly matches between all-reduce & reduce-scatter + all-gather**, which means each layer is exactly same with no accumulated error. So I wonder
1. Can you tell me where there is two dummy run. in My example of Qwen3-32B, one seqlen is max model len, one seqlen is 1024
2. Can you possibly tell me What may influence the precision ?
### How would you like to use vllm
_No response_
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/31361 | closed | [
"usage"
] | 2025-12-25T16:38:03Z | 2025-12-27T03:41:27Z | 0 | Dingjifeng |
vllm-project/vllm | 31,353 | [Bug]: KV Cache grows continuously with just one chat completion request using meta-llama/Llama-3.2-1B on L40 GPU with Flash Attention and finally completed after 10 minutes | ### Your current environment
<details>
<summary>The output of <code>python collect_env.py</code></summary>
```text
Collecting environment information...
==============================
System Info
==============================
OS : Ubuntu 24.04.3 LTS (x86_64)
GCC version : (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
Clang version : Could not collect
CMake version : version 3.28.3
Libc version : glibc-2.39
==============================
PyTorch Info
==============================
PyTorch version : 2.9.0+cu128
Is debug build : False
CUDA used to build PyTorch : 12.8
ROCM used to build PyTorch : N/A
==============================
Python Environment
==============================
Python version : 3.12.11 | packaged by Anaconda, Inc. | (main, Jun 5 2025, 13:09:17) [GCC 11.2.0] (64-bit runtime)
Python platform : Linux-5.15.0-161-generic-x86_64-with-glibc2.39
==============================
CUDA / GPU Info
==============================
Is CUDA available : True
CUDA runtime version : 12.6.85
CUDA_MODULE_LOADING set to :
GPU models and configuration : GPU 0: NVIDIA L40S
Nvidia driver version : 550.163.01
cuDNN version : Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.9.5.1
/usr/lib/x86_64-linux-gnu/libcudnn_adv.so.9.5.1
/usr/lib/x86_64-linux-gnu/libcudnn_cnn.so.9.5.1
/usr/lib/x86_64-linux-gnu/libcudnn_engines_precompiled.so.9.5.1
/usr/lib/x86_64-linux-gnu/libcudnn_engines_runtime_compiled.so.9.5.1
/usr/lib/x86_64-linux-gnu/libcudnn_graph.so.9.5.1
/usr/lib/x86_64-linux-gnu/libcudnn_heuristic.so.9.5.1
/usr/lib/x86_64-linux-gnu/libcudnn_ops.so.9.5.1
HIP runtime version : N/A
MIOpen runtime version : N/A
Is XNNPACK available : True
==============================
CPU Info
==============================
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 46 bits physical, 57 bits virtual
Byte Order: Little Endian
CPU(s): 16
On-line CPU(s) list: 0-15
Vendor ID: GenuineIntel
Model name: Intel(R) Xeon(R) Gold 6338 CPU @ 2.00GHz
CPU family: 6
Model: 106
Thread(s) per core: 2
Core(s) per socket: 8
Socket(s): 1
Stepping: 6
BogoMIPS: 3990.65
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid tsc_known_freq pni pclmulqdq dtes64 ssse3 fma cx16 pdcm pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch cpuid_fault invpcid_single ssbd ibrs ibpb stibp ibrs_enhanced fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves wbnoinvd arat avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid fsrm md_clear arch_capabilities
Hypervisor vendor: KVM
Virtualization type: full
L1d cache: 512 KiB (16 instances)
L1i cache: 512 KiB (16 instances)
L2 cache: 32 MiB (8 instances)
L3 cache: 16 MiB (1 instance)
NUMA node(s): 1
NUMA node0 CPU(s): 0-15
Vulnerability Gather data sampling: Unknown: Dependent on hypervisor status
Vulnerability Indirect target selection: Mitigation; Aligned branch/return thunks
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Vulnerable: Clear CPU buffers attempted, no microcode; SMT Host state unknown
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed: Not affected
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Enhanced / Automatic IBRS; IBPB conditional; | https://github.com/vllm-project/vllm/issues/31353 | open | [
"bug",
"help wanted"
] | 2025-12-25T13:56:52Z | 2025-12-27T15:55:34Z | 1 | aravilli |
sgl-project/sglang | 15,825 | Is it normal that Qwen3-30B-A3B runs slower than Qwen3-8B? | I serve two models on the Ascend 910 platform (following sglang's ascend examples) with the same tp2dp8 and benchmarked them.
Before testing, I suppose A3B will be faster than 8B for fewer activated tensor blocks.
But the result is different:
### qwen 30B A3B
```
export SGLANG_SET_CPU_AFFINITY=1
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export STREAMS_PER_DEVICE=32
export HCCL_BUFFSIZE=1536
export HCCL_OP_EXPANSION_MODE=AIV
export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=32
export SGLANG_DEEPEP_BF16_DISPATCH=1
export ENABLE_ASCEND_MOE_NZ=1
python -m sglang.launch_server \
--device npu \
--attention-backend ascend \
--trust-remote-code \
--tp-size 2 \
--dp-size 8 \
--model **Qwen/Qwen3-30B-A3B-Instruct-2507** \
--model-path /models/Qwen3-30B-A3B-Instruct-2507 \
--port 30111 \
--mem-fraction-static 0.8
```
```
============ Serving Benchmark Result ============
Backend: sglang
Traffic request rate: inf
Max request concurrency: not set
Successful requests: 1000
Benchmark duration (s): 69.68
Total input tokens: 3055233
Total input text tokens: 3055233
Total input vision tokens: 0
Total generated tokens: 513413
Total generated tokens (retokenized): 512578
Request throughput (req/s): 14.35
Input token throughput (tok/s): 43846.56
**Output token throughput (tok/s): 7368.14**
Peak output token throughput (tok/s): 12775.00
Peak concurrent requests: 1000
Total token throughput (tok/s): 51214.70
Concurrency: 665.97
----------------End-to-End Latency----------------
Mean E2E Latency (ms): 46404.83
Median E2E Latency (ms): 49605.93
---------------Time to First Token----------------
Mean TTFT (ms): 10682.85
Median TTFT (ms): 9808.31
P99 TTFT (ms): 16320.45
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms): 96.14
Median TPOT (ms): 75.08
P99 TPOT (ms): 399.24
---------------Inter-Token Latency----------------
Mean ITL (ms): 69.71
Median ITL (ms): 69.43
P95 ITL (ms): 80.73
P99 ITL (ms): 96.53
Max ITL (ms): 5450.67
==================================================
```
### Qwen3 8B
```
export SGLANG_SET_CPU_AFFINITY=1
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export STREAMS_PER_DEVICE=32
export HCCL_BUFFSIZE=1536
export HCCL_OP_EXPANSION_MODE=AIV
ASCEND_RT_VISIBLE_DEVICES=0 python -m sglang.launch_server \
--device npu \
--attention-backend ascend \
--trust-remote-code \
--model Qwen/Qwen3-8B \
--model-path /models/Qwen3-8B \
--port 30111 \
--mem-fraction-static 0.8 \
--tp-size 2 \
--dp-size 8
```
```
============ Serving Benchmark Result ============
Backend: sglang
Traffic request rate: inf
Max request concurrency: not set
Successful requests: 1000
Benchmark duration (s): 49.67
Total input tokens: 3055233
Total input text tokens: 3055233
Total input vision tokens: 0
Total generated tokens: 513413
Total generated tokens (retokenized): 512976
Request throughput (req/s): 20.13
Input token throughput (tok/s): 61513.14
**Output token throughput (tok/s): 10336.90**
Peak output token throughput (tok/s): 23242.00
Peak concurrent requests: 1000
Total token throughput (tok/s): 71850.04
Concurrency: 709.69
----------------End-to-End Latency----------------
Mean E2E Latency (ms): 35249.04
Median E2E Latency (ms): 36490.95
---------------Time to First Token----------------
Mean TTFT (ms): 10977.22
Median TTFT (ms): 9339.57
P99 TTFT (ms): 16697.36
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms): 82.35
Median TPOT (ms): 48.71
P99 TPOT (ms): 516.74
---------------Inter-Token Latency----------------
Mean ITL (ms): 47.37
Median ITL (ms): 35.12
P95 ITL (ms): 105.74
P99 ITL (ms): 463.46
Max I | https://github.com/sgl-project/sglang/issues/15825 | open | [] | 2025-12-25T11:26:10Z | 2025-12-25T11:26:10Z | 0 | yucc-leon |
vllm-project/vllm | 31,344 | [Usage]: how to pass param logits_processors in AsyncEngineArgs? | ### Your current environment
import torch
from transformers import LogitsProcessor
from transformers.generation.logits_process import _calc_banned_ngram_tokens
from typing import List, Set
class NoRepeatNGramLogitsProcessor(LogitsProcessor):
def __init__(self, ngram_size: int, window_size: int = 100, whitelist_token_ids: set = None):
if not isinstance(ngram_size, int) or ngram_size <= 0:
raise ValueError(f"`ngram_size` has to be a strictly positive integer, but is {ngram_size}")
if not isinstance(window_size, int) or window_size <= 0:
raise ValueError(f"`window_size` has to be a strictly positive integer, but is {window_size}")
self.ngram_size = ngram_size
self.window_size = window_size
self.whitelist_token_ids = whitelist_token_ids or set()
def __call__(self, input_ids: List[int], scores: torch.FloatTensor) -> torch.FloatTensor:
if len(input_ids) < self.ngram_size:
return scores
current_prefix = tuple(input_ids[-(self.ngram_size - 1):])
search_start = max(0, len(input_ids) - self.window_size)
search_end = len(input_ids) - self.ngram_size + 1
banned_tokens = set()
for i in range(search_start, search_end):
ngram = tuple(input_ids[i:i + self.ngram_size])
if ngram[:-1] == current_prefix:
banned_tokens.add(ngram[-1])
banned_tokens = banned_tokens - self.whitelist_token_ids
if banned_tokens:
scores = scores.clone()
for token in banned_tokens:
scores[token] = -float("inf")
return scores
async def stream_generate(image=None, prompt=''):
logits_processors = [NoRepeatNGramLogitsProcessor(ngram_size=30, window_size=90,
whitelist_token_ids={128821, 128822})] # whitelist: <td>, </td>
#高版本
logits_processors_config: list[Dict[str, Any]] = [
{
"class": NoRepeatNGramLogitsProcessor, # 传入类对象
"kwargs": { # 初始化参数
"ngram_size": 30,
"window_size": 90,
"whitelist_token_ids": {128821, 128822}
}
}
]
engine_args = AsyncEngineArgs(
model=MODEL_PATH,
#hf_overrides={"architectures": ["DeepseekOCRForCausalLM"]},
block_size=256,
max_model_len=8192,
enforce_eager=False,
trust_remote_code=True,
tensor_parallel_size=1,
gpu_memory_utilization=0.75,
logits_processors=logits_processors_config
)
engine = AsyncLLMEngine.from_engine_args(engine_args)
error:
".local/lib/python3.13/site-packages/vllm/engine/arg_utils.py", line 1189, in create_model_config
return ModelConfig(
model=self.model,
...<46 lines>...
io_processor_plugin=self.io_processor_plugin,
)
File "/.local/lib/python3.13/site-packages/pydantic/_internal/_dataclasses.py", line 121, in __init__
s.__pydantic_validator__.validate_python(ArgsKwargs(args, kwargs), self_instance=s)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
pydantic_core._pydantic_core.ValidationError: 2 validation errors for ModelConfig
logits_processors.0.str
Input should be a valid string [type=string_type, input_value={'class': <class 'process...ids': {128821, 128822}}}, input_type=dict]
For further information visit https://errors.pydantic.dev/2.12/v/string_type
logits_processors.0.custom-error[is-instance[type]]
Input should be a type [type=is_type, input_value={'class': <class 'process...ids': {128821, 128822}}}, input_type=dict] | https://github.com/vllm-project/vllm/issues/31344 | open | [
"usage"
] | 2025-12-25T10:12:02Z | 2025-12-25T13:30:54Z | 0 | cqray1990 |
huggingface/diffusers | 12,889 | Question about qwen-image-edit-2511 loading warning | When loading the model qwen-image-edit-2511 using the diffusers library, I encounter the following warning:
The config attributes {'zero_cond_t': True} were passed to QwenImageTransformer2DModel, but are not expected and will be ignored. Please verify your config.json configuration file.
This suggests that the zero_cond_t parameter is present in the model’s config but is not recognized by the current implementation of QwenImageTransformer2DModel. Could you please clarify whether this attribute is deprecated, optional, or requires a specific version of the library? Additionally, is there any recommended action to suppress or resolve this warning? | https://github.com/huggingface/diffusers/issues/12889 | closed | [] | 2025-12-25T07:06:28Z | 2025-12-25T08:56:28Z | 2 | wizardbob |
sgl-project/sglang | 15,810 | [Bug] hicache 3fs backend global metadata much instance deploy bug | ### Checklist
- [x] I searched related issues but found no solution.
- [x] The bug persists in the latest version.
- [ ] Issues without environment info and a minimal reproducible demo are hard to resolve and may receive no feedback.
- [ ] If this is not a bug report but a general question, please start a discussion at https://github.com/sgl-project/sglang/discussions. Otherwise, it will be closed.
- [ ] Please use English. Otherwise, it will be closed.
### Describe the bug
Currently, although the 3fs backend uses the globalMetadata service to manage the global 3fs cache, the following issue exists: this service cannot be deployed with multiple instances. Multi-instance deployment would cause concurrent write problems with metadata. Is there a chance to fix this issue?
### Reproduction
check code can know this
### Environment
。。 | https://github.com/sgl-project/sglang/issues/15810 | open | [] | 2025-12-25T06:52:45Z | 2025-12-25T09:42:30Z | 4 | weibingo |
vllm-project/vllm | 31,319 | [Bug]: GLM-4.7-FP8 missing beginning <think> tag | ### Your current environment
I am on docker nightly vLLM API server version 0.14.0rc1.dev104+g8ee90c83f
### 🐛 Describe the bug
I hosted the model via vllm and already without reasoning_parser, I found the model output with directly output without <think> but having close tag </think> later.
```
root@iv-ydzbs5zshss6ipm6s5gu /h/n/d/ark_http_proxy# curl --location 'http://localhost/v1/chat/completions' \
--header 'Authorization: Bearer YOUR_API_KEY' \
--header 'Content-Type: application/json' \
--data '{
"model": "GLM-4.7-FP8", "stream": true,
"messages": [
{
"role": "user",
"content": "what is cryptography"
}
],"chat_template_kwargs": {"enable_thinking": true}, "skip_special_tokens": false,
"thinking": {
"type": "enabled"
},
"max_tokens": 1024,
"temperature": 1.0
}'
data: {"id":"chatcmpl-9fbc092d919f9e51","object":"chat.completion.chunk","created":1766599479,"model":"GLM-4.7-FP8","choices":[{"index":0,"delta":{"role":"assistant","content":"","reasoning_content":null},"logprobs":null,"finish_reason":null}],"prompt_token_ids":null}
data: {"id":"chatcmpl-9fbc092d919f9e51","object":"chat.completion.chunk","created":1766599479,"model":"GLM-4.7-FP8","choices":[{"index":0,"delta":{"content":"1","reasoning_content":null},"logprobs":null,"finish_reason":null,"token_ids":null}]}
data: {"id":"chatcmpl-9fbc092d919f9e51","object":"chat.completion.chunk","created":1766599479,"model":"GLM-4.7-FP8","choices":[{"index":0,"delta":{"content":". ","reasoning_content":null},"logprobs":null,"finish_reason":null,"token_ids":null}]}
data: {"id":"chatcmpl-9fbc092d919f9e51","object":"chat.completion.chunk","created":1766599479,"model":"GLM-4.7-FP8","choices":[{"index":0,"delta":{"content":" **An","reasoning_content":null},"logprobs":null,"finish_reason":null,"token_ids":null}]}
data: {"id":"chatcmpl-9fbc092d919f9e51","object":"chat.completion.chunk","created":1766599479,"model":"GLM-4.7-FP8","choices":[{"index":0,"delta":{"content":"alyze the","reasoning_content":null},"logprobs":null,"finish_reason":null,"token_ids":null}]}
```
I confirmed that chat template will
```
root@iv-ydzbs5zshss6ipm6s5gu /h/n/d/ark_http_proxy# curl -sS 'http://127.0.0.1/tokenize' \
-H 'Content-Type: application/json' \
-d '{"model":"GLM-4.7-FP8","messages":[{"role":"user","content":"hi"}],"add_generation_prompt":true,"return_token_strs":true}'
{"count":6,"max_model_len":202752,"tokens":[151331,151333,151336,6023,151337,151350],"token_strs":["[gMASK]","<sop>","<|user|>","hi","<|assistant|>","<think>"]}⏎
```
I think we need a similar **minimax_m2_append_think** reasoning parser to simply append think to content beginning?
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/31319 | open | [
"bug"
] | 2025-12-24T18:45:34Z | 2026-01-06T07:59:45Z | 16 | Nemo-G |
vllm-project/vllm | 31,278 | [Usage]:请问Qwen3-VL本地加载模式支持单独加载LoRA么？ | 请问Qwen3-VL本地加载模式支持单独加载LoRA么？ | https://github.com/vllm-project/vllm/issues/31278 | open | [
"usage"
] | 2025-12-24T11:33:08Z | 2025-12-25T03:52:16Z | 3 | dengdeng-cat |
vllm-project/vllm | 31,272 | [Performance]: b200x8 deepseek-ai/DeepSeek-V3.2-Exp max perf | ### Proposal to improve performance
_No response_
### Report of performance regression
Do you have any ideas on how to increase TPS? I have two servers — one with H200 ×8 and another with B200 ×8. They use the same startup script, but the performance is almost identical. In my opinion, B200 should be faster than H200, so maybe my settings are not optimal
vllm serve \
--model deepseek-ai/DeepSeek-V3.2-Exp \
--served-model-name deepseek-ai/DeepSeek-V3.2-Exp \
--host 0.0.0.0 \
--port 12345 \
--tensor-parallel-size 8 \
--enable-auto-tool-choice \
--tool-call-parser deepseek_v31 \
--chat-template /root/tool_chat_template_deepseekv31.jinja \
--gpu-memory-utilization 0.9 \
--max-model-len 125000 \
### Misc discussion on performance
_No response_
### Your current environment (if you think it is necessary)
```text
Collecting environment information...
uv is set
==============================
System Info
==============================
OS : Ubuntu 24.04.3 LTS (x86_64)
GCC version : (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
Clang version : Could not collect
CMake version : Could not collect
Libc version : glibc-2.39
==============================
PyTorch Info
==============================
PyTorch version : 2.9.0+cu128
Is debug build : False
CUDA used to build PyTorch : 12.8
ROCM used to build PyTorch : N/A
==============================
Python Environment
==============================
Python version : 3.12.3 (main, Nov 6 2025, 13:44:16) [GCC 13.3.0] (64-bit runtime)
Python platform : Linux-6.8.0-87-generic-x86_64-with-glibc2.39
==============================
CUDA / GPU Info
==============================
Is CUDA available : True
CUDA runtime version : 13.0.88
CUDA_MODULE_LOADING set to :
GPU models and configuration :
GPU 0: NVIDIA B200
GPU 1: NVIDIA B200
GPU 2: NVIDIA B200
GPU 3: NVIDIA B200
GPU 4: NVIDIA B200
GPU 5: NVIDIA B200
GPU 6: NVIDIA B200
GPU 7: NVIDIA B200
Nvidia driver version : 580.95.05
cuDNN version : Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.9.14.0
/usr/lib/x86_64-linux-gnu/libcudnn_adv.so.9.14.0
/usr/lib/x86_64-linux-gnu/libcudnn_cnn.so.9.14.0
/usr/lib/x86_64-linux-gnu/libcudnn_engines_precompiled.so.9.14.0
/usr/lib/x86_64-linux-gnu/libcudnn_engines_runtime_compiled.so.9.14.0
/usr/lib/x86_64-linux-gnu/libcudnn_graph.so.9.14.0
/usr/lib/x86_64-linux-gnu/libcudnn_heuristic.so.9.14.0
/usr/lib/x86_64-linux-gnu/libcudnn_ops.so.9.14.0
HIP runtime version : N/A
MIOpen runtime version : N/A
Is XNNPACK available : True
==============================
CPU Info
==============================
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 52 bits physical, 57 bits virtual
Byte Order: Little Endian
CPU(s): 240
On-line CPU(s) list: 0-239
Vendor ID: AuthenticAMD
BIOS Vendor ID: QEMU
Model name: AMD EPYC 9575F 64-Core Processor
BIOS Model name: pc-q35-8.2 CPU @ 2.0GHz
BIOS CPU family: 1
CPU family: 26
Model: 2
Thread(s) per core: 1
Core(s) per socket: 1
Socket(s): 240
Stepping: 1
BogoMIPS: 6590.10
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm rep_good nopl cpuid extd_apicid tsc_known_freq pni pclmulqdq ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm cmp_legacy svm cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw perfctr_core ssbd ibrs ibpb stibp ibrs_enhanced vmmcall fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves avx_vnni avx512_bf16 clzero xsaveerptr wbnoinvd arat npt lbrv nrip_save tsc_scale vmcb_clean flushbyasid pausefilter pfthreshold v_vmsave_vmload vgif vnmi avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid movdiri movdir64b fsrm avx512_vp2intersect flush_l1d arch_capabilities
Virtualization: AMD-V
Hypervisor vendor: KVM
Virtualization type: full
L1d cache: 15 MiB (240 instances)
L1i cache: 15 MiB (240 instances)
L2 cache: | https://github.com/vllm-project/vllm/issues/31272 | open | [
"performance"
] | 2025-12-24T09:48:01Z | 2025-12-24T10:09:29Z | 0 | evgeniiperepelkin |
huggingface/trl | 4,747 | Addition of Supervised Reinforcement Learning | ### Feature request
https://arxiv.org/pdf/2510.25992 can i work on its implementation ?
### Motivation
Better approach then previous RL's
### Your contribution
I can work on it following reference paper | https://github.com/huggingface/trl/issues/4747 | open | [] | 2025-12-24T09:20:32Z | 2025-12-24T09:20:32Z | 0 | kushalgarg101 |
vllm-project/vllm | 31,270 | [Bug]: Can run Speculative decode with PP >2? | ### Your current environment
vllm:0.12.0
### 🐛 Describe the bug
I run vllm:0.12.0 with start args like this:
`python3 -m vllm.entrypoints.openai.api_server \
--host 0.0.0.0 --port 8080 --dtype bfloat16 --model /Qwen3-32B \
--pipeline-parallel-size 2 \
--gpu-memory-utilization 0.9 --max-model-len 32768 --max-num-batched-tokens 5120 \
--trust-remote-code --no-enable-prefix-caching \
--speculative_config '{"method": "ngram","num_speculative_tokens": 10,"prompt_lookup_max": 4, "enforce_eager": "True"}'`
The server can start, but when use the interface of '/chat/completion', the vllm server will crash.
### Before submitting a new issue...
- [ ] #31271 | https://github.com/vllm-project/vllm/issues/31270 | open | [
"bug"
] | 2025-12-24T09:10:05Z | 2025-12-26T07:27:11Z | 1 | frankie-ys |
sgl-project/sglang | 15,739 | [Bug] Failed to deploy DeepSeek-V3.2 with LMCache | ### Checklist
- [x] I searched related issues but found no solution.
- [x] The bug persists in the latest version.
- [x] Issues without environment info and a minimal reproducible demo are hard to resolve and may receive no feedback.
- [x] If this is not a bug report but a general question, please start a discussion at https://github.com/sgl-project/sglang/discussions. Otherwise, it will be closed.
- [x] Please use English. Otherwise, it will be closed.
### Describe the bug
I use v0.5.6.post2 with LMCache 0.3.10 to deploy DeepSeek-V3.2.
I got the following error :
```
[2025-12-24 08:20:12 PP0 TP2 EP2] Scheduler hit an exception: Traceback (most recent call last):
File "/sgl-workspace/sglang/python/sglang/srt/managers/scheduler.py", line 2680, in run_scheduler_process
scheduler = Scheduler(
^^^^^^^^^^
File "/sgl-workspace/sglang/python/sglang/srt/managers/scheduler.py", line 434, in __init__
self.init_cache_with_memory_pool()
File "/sgl-workspace/sglang/python/sglang/srt/managers/scheduler.py", line 781, in init_cache_with_memory_pool
self.tree_cache = LMCRadixCache(
^^^^^^^^^^^^^^
File "/sgl-workspace/sglang/python/sglang/srt/mem_cache/storage/lmcache/lmc_radix_cache.py", line 91, in __init__
getattr(self.token_to_kv_pool_allocator._kvcache, "k_buffer"),
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: 'NSATokenToKVPool' object has no attribute 'k_buffer'. Did you mean: 'kv_buffer'?
```
Is there anything wrong with my configuration? Please advise.
Thanks~
### Reproduction
My configs:
> lmcache_config.yaml
```
chunk_size: 256
local_cpu: true
max_local_cpu_size: 5.0
#
remote_url: "redis://10.62.207.53:32628"
remote_serde: "naive"
```
> master.sh
```
export LMCACHE_CONFIG_PATH=/mnt/scripts/lmcache_config.yaml
export LMCACHE_ENABLE=True
python -m sglang.launch_server \
--model-path=/mnt/models/deepseek-ai/DeepSeek-V3.2 \
--served-model-name=deepseek-ai/DeepSeek-V3.2 \
--tensor-parallel-size=4 \
--pipeline-parallel-size=2 \
--expert-parallel-size=4 \
--data-parallel-size=1 \
--enable-dp-attention \
--trust-remote-code \
--mem-fraction-static=0.8 \
--log-requests \
--log-requests-level=3 \
--dist-init-addr="${MASTER_IP}:${PORT}" \
--nnodes="$NNODES" \
--node-rank="$NODE_RANK" \
--tool-call-parser=deepseekv32 \
--reasoning-parser=deepseek-v3 \
--host=0.0.0.0 \
--port=8000 \
--enable-lmcache \
--enable-metrics
```
> worker.sh
```
export LMCACHE_CONFIG_PATH=/mnt/scripts/lmcache_config.yaml
export LMCACHE_ENABLE=True
python -m sglang.launch_server \
--model-path=/mnt/models/deepseek-ai/DeepSeek-V3.2 \
--served-model-name=deepseek-ai/DeepSeek-V3.2 \
--tensor-parallel-size=4 \
--pipeline-parallel-size=2 \
--expert-parallel-size=4 \
--data-parallel-size=1 \
--enable-dp-attention \
--trust-remote-code \
--mem-fraction-static=0.8 \
--log-requests \
--log-requests-level=3 \
--dist-init-addr="${MASTER_IP}:${PORT}" \
--nnodes="$NNODES" \
--node-rank="$NODE_RANK" \
--tool-call-parser=deepseekv32 \
--reasoning-parser=deepseek-v3 \
--enable-lmcache \
--enable-metrics
```
### Environment
sglang: v0.5.6.post2
lmcache: v0.3.10
model: DeepSeek-V3.2 | https://github.com/sgl-project/sglang/issues/15739 | open | [] | 2025-12-24T08:45:29Z | 2025-12-29T22:55:27Z | 1 | niceallen |
sgl-project/sglang | 15,710 | [Bug] Using TBO, but no overlap in decoding phase? | ### Checklist
- [x] I searched related issues but found no solution.
- [x] The bug persists in the latest version.
- [x] Issues without environment info and a minimal reproducible demo are hard to resolve and may receive no feedback.
- [x] If this is not a bug report but a general question, please start a discussion at https://github.com/sgl-project/sglang/discussions. Otherwise, it will be closed.
- [x] Please use English. Otherwise, it will be closed.
### Describe the bug
<!-- Failed to upload "cf3c6fa8b605e6bbe3cb65ceee9bd06.png" -->
### Reproduction
python -m sglang.launch_server --model-path /root/temp_can/DeepSeek-V3-0324 --load-format dummy --tp 4 --ep 4 --moe-a2a-backend deepep --deepep-mode auto --chunked-prefill-size -1 --host 0.0.0.0 --port 30000 --enable-two-batch-overlap --mem-fraction-static 0.4
python3 -m sglang.bench_one_batch_server --model-path /root/temp_can/DeepSeek-V3-0324 --base-url http://127.0.0.1:30000 --batch-size 256 --input-len 64 --output-len 128 --skip-warmup --profile
### Environment
(new_py310) root@zyhuang0-0:~/temp_can/sglang# python3 -m sglang.check_env
Python: 3.10.19 (main, Oct 21 2025, 16:43:05) [GCC 11.2.0]
CUDA available: True
GPU 0,1: NVIDIA H100 80GB HBM3
GPU 0,1 Compute Capability: 9.0
CUDA_HOME: /usr/local/cuda
NVCC: Cuda compilation tools, release 12.9, V12.9.41
CUDA Driver Version: 550.54.15
PyTorch: 2.9.1+cu128
sglang: 0.5.6.post2
sgl_kernel: 0.3.19
flashinfer_python: 0.5.3
flashinfer_cubin: 0.5.3
flashinfer_jit_cache: Module Not Found
triton: 3.5.1
transformers: 4.57.1
torchao: 0.9.0
numpy: 2.2.6
aiohttp: 3.13.2
fastapi: 0.127.0
hf_transfer: 0.1.9
huggingface_hub: 0.36.0
interegular: 0.3.3
modelscope: 1.33.0
orjson: 3.11.5
outlines: 0.1.11
packaging: 25.0
psutil: 7.1.3
pydantic: 2.12.5
python-multipart: 0.0.21
pyzmq: 27.1.0
uvicorn: 0.40.0
uvloop: 0.22.1
vllm: Module Not Found
xgrammar: 0.1.27
openai: 2.6.1
tiktoken: 0.12.0
anthropic: 0.75.0
litellm: Module Not Found
decord2: 3.0.0
NVIDIA Topology:
GPU0 GPU1 NIC0 NIC1 NIC2 NIC3 NIC4 NIC5 NIC6 CPU Affinity NUMA Affinity GPU NUMA ID
GPU0 X NV18 SYS PIX SYS SYS SYS SYS SYS 0-47,96-143 0 N/A
GPU1 NV18 X SYS SYS SYS SYS SYS PIX SYS 48-95,144-191 1 N/A
NIC0 SYS SYS X SYS SYS SYS SYS SYS SYS
NIC1 PIX SYS SYS X SYS SYS SYS SYS SYS
NIC2 SYS SYS SYS SYS X PXB PXB SYS SYS
NIC3 SYS SYS SYS SYS PXB X PIX SYS SYS
NIC4 SYS SYS SYS SYS PXB PIX X SYS SYS
NIC5 SYS PIX SYS SYS SYS SYS SYS X SYS
NIC6 SYS SYS SYS SYS SYS SYS SYS SYS X
Legend:
X = Self
SYS = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
PHB = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
PXB = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
PIX = Connection traversing at most a single PCIe bridge
NV# = Connection traversing a bonded set of # NVLinks
NIC Legend:
NIC0: mlx5_0
NIC1: mlx5_1
NIC2: mlx5_2
NIC3: mlx5_3
NIC4: mlx5_4
NIC5: mlx5_5
NIC6: mlx5_6
ulimit soft: 1048576 | https://github.com/sgl-project/sglang/issues/15710 | open | [] | 2025-12-24T02:22:19Z | 2025-12-24T02:22:19Z | 0 | ziyuhuang123 |
sgl-project/sglang | 15,707 | [Feature] diffusion: TurboDiffusion achieves a 200x speedup on a single GPU, bringing video into the second-level era | ### Checklist
- [ ] If this is not a feature request but a general question, please start a discussion at https://github.com/sgl-project/sglang/discussions. Otherwise, it will be closed.
- [ ] Please use English. Otherwise, it will be closed.
### Motivation
https://github.com/thu-ml/TurboDiffusion
When can it be integrated into sglang-diffusion ?
> [清华系 DeepSeek 时刻来了，硅谷沸腾！单卡 200 倍加速，视频进入秒级时代](https://mp.weixin.qq.com/s/JmHwMsCYr9M39JLy1jAb7A)
### Related resources
_No response_ | https://github.com/sgl-project/sglang/issues/15707 | open | [] | 2025-12-24T01:50:02Z | 2025-12-30T08:45:43Z | 1 | xiaolin8 |
huggingface/transformers | 43,023 | How to investigate "CAS service error" during model downloading? | ### System Info
(nm) PS C:\Users\myuser\AppData\Local\anaconda3\envs\nm\Lib\site-packages\transformers\commands> python .\transformers_cli.py env
```
Copy-and-paste the text below in your GitHub issue and FILL OUT the two last points.
- `transformers` version: 4.57.3
- Platform: Windows-10-10.0.19045-SP0
- Python version: 3.10.19
- Huggingface_hub version: 0.36.0
- Safetensors version: 0.7.0
- Accelerate version: not installed
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch version (accelerator?): 2.7.0 (NA)
- Tensorflow version (GPU?): not installed (NA)
- Flax version (CPU?/GPU?/TPU?): not installed (NA)
- Jax version: not installed
- JaxLib version: not installed
- Using distributed or parallel set-up in script?: the whole code posted below
```
### Who can help?
_No response_
### Information
- [ ] The official example scripts
- [x] My own modified scripts
### Tasks
- [ ] An officially supported task in the `examples` folder (such as GLUE/SQuAD, ...)
- [ ] My own task or dataset (give details below)
### Reproduction
Base example from [here](https://huggingface.co/cross-encoder/ms-marco-MiniLM-L6-v2)
```
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
model = AutoModelForSequenceClassification.from_pretrained('cross-encoder/ms-marco-MiniLM-L6-v2')
tokenizer = AutoTokenizer.from_pretrained('cross-encoder/ms-marco-MiniLM-L6-v2')
features = tokenizer(['How many people live in Berlin?', 'How many people live in Berlin?'], ['Berlin has a population of 3,520,031 registered inhabitants in an area of 891.82 square kilometers.', 'New York City is famous for the Metropolitan Museum of Art.'], padding=True, truncation=True, return_tensors="pt")
model.eval()
with torch.no_grad():
scores = model(**features).logits
print(scores)
```
returns
```
model.safetensors: 0%
0.00/90.9M [00:32<?, ?B/s]
---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
File c:\Users\myuser\AppData\Local\anaconda3\envs\nm\lib\site-packages\transformers\modeling_utils.py:1037, in _get_resolved_checkpoint_files(pretrained_model_name_or_path, subfolder, variant, gguf_file, from_tf, from_flax, use_safetensors, cache_dir, force_download, proxies, local_files_only, token, user_agent, revision, commit_hash, is_remote_code, transformers_explicit_filename)
1024 cached_file_kwargs = {
1025 "cache_dir": cache_dir,
1026 "force_download": force_download,
(...)
1035 "_commit_hash": commit_hash,
1036 }
-> [1037](file:///C:/Users/myuser /AppData/Local/anaconda3/envs/nm/lib/site-packages/transformers/modeling_utils.py:1037) resolved_archive_file = cached_file(pretrained_model_name_or_path, filename, **cached_file_kwargs)
1039 # Since we set _raise_exceptions_for_missing_entries=False, we don't get an exception but a None
1040 # result when internet is up, the repo and revision exist, but the file does not.
File c:\Users\myuser\AppData\Local\anaconda3\envs\nm\lib\site-packages\transformers\utils\hub.py:322, in cached_file(path_or_repo_id, filename, **kwargs)
269 """
270 Tries to locate a file in a local folder and repo, downloads and cache it if necessary.
271
(...)
320 ```
321 """
--> [322](file:///C:/Users/myuser /AppData/Local/anaconda3/envs/nm/lib/site-packages/transformers/utils/hub.py:322) file = cached_files(path_or_repo_id=path_or_repo_id, filenames=[filename], **kwargs)
323 file = file[0] if file is not None else file
File c:\Users\myuser\AppData\Local\anaconda3\envs\nm\lib\site-packages\transformers\utils\hub.py:567, in cached_files(path_or_repo_id, filenames, cache_dir, force_download, resume_download, proxies, token, revision, local_files_only, subfolder, repo_type, user_agent, _raise_exceptions_for_gated_repo, _raise_exceptions_for_missing_entries, _raise_exceptions_for_connection_errors, _commit_hash, **deprecated_kwargs)
566 elif not isinstance(e, EntryNotFoundError):
--> [567](file:///C:/Users/myuser /AppData/Local/anaconda3/envs/nm/lib/site-packages/transformers/utils/hub.py:567) raise e
569 resolved_files = [
570 _get_cache_file_to_return(path_or_repo_id, filename, cache_dir, revision) for filename in full_filenames
571 ]
File c:\Users\myuser\AppData\Local\anaconda3\envs\nm\lib\site-packages\transformers\utils\hub.py:479, in cached_files(path_or_repo_id, filenames, cache_dir, force_download, resume_download, proxies, token, revision, local_files_only, subfolder, repo_type, user_agent, _raise_exceptions_for_gated_repo, _raise_exceptions_for_missing_entries, _raise_exceptions_for_connection_errors, _commit_hash, **deprecated_kwargs)
477 if len(full_filenames) == 1:
478 # This is slightly better for only 1 file
--> [479](file:///C:/Users/myuser /AppData/Local/anaconda3/envs/nm/lib/site-packages/transformers/utils/hub.py:479) hf_hub_download | https://github.com/huggingface/transformers/issues/43023 | open | [
"bug"
] | 2025-12-23T14:48:51Z | 2025-12-25T14:36:42Z | null | satyrmipt |
vllm-project/vllm | 31,217 | [Usage]: suffix decoding | ### Your current environment
Does suffix decoding necessarily require a repetition penalty of 1?
### How would you like to use vllm
Does suffix decoding necessarily require a repetition penalty of 1?
In suffix decoding, I found that when the repetition penalty is not equal to 1, the acceleration is not significant. However, when the repetition penalty is equal to 1, the acceleration is very noticeable.
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/31217 | open | [
"usage"
] | 2025-12-23T10:43:45Z | 2025-12-24T02:56:35Z | 1 | jiangix-paper |
huggingface/lerobot | 2,707 | Transformers dependency | ### Ticket Type
🐛 Bug Report (Something isn't working)
### Environment & System Info
```Shell
- lerobot version: 0.4.3
- Platform: Linux-5.14.0-570.26.1.el9_6.x86_64-x86_64-with-glibc2.34
- Python version: 3.12.12
- Huggingface Hub version: 0.35.3
- Datasets version: 4.1.1
- Numpy version: 2.3.5
- PyTorch version: 2.7.1
- Is PyTorch built with CUDA support?: False
- Cuda version: N/A
- GPU model: N/A
- Using GPU in script?: <fill in>
```
### Description
Hi,
Since commit f04958527e70cac3aa95265badd97b53f3ef7633 and the dependency bump of transformers from 4.53 to 4.57, some extras are conflicting (pi that requires a custom 4.53 and almost all others that requires >=4.57), so I can't upgrade. Was the bump really necessary ? Or is there a possibility to get rid of the custom openpi transformers ?
### Context & Reproduction
[pypi-dependencies]
lerobot = {git = "https://github.com/huggingface/lerobot.git", extras = ["smolvla", "pi", "groot"]}
pixi install
### Relevant logs or stack trace
```Shell
```
### Checklist
- [x] I have searched existing tickets to ensure this isn't a duplicate.
- [x] I am using the latest version of the `main` branch.
- [x] I have verified this is not an environment-specific problem.
### Additional Info / Workarounds
_No response_ | https://github.com/huggingface/lerobot/issues/2707 | closed | [
"bug",
"question",
"dependencies"
] | 2025-12-23T10:37:53Z | 2025-12-23T23:43:10Z | null | RomDeffayet |
vllm-project/vllm | 31,216 | [RFC]: Sampling Optimization: move gather of logits after argmax. | ### Motivation.
As shown in the left part of the following picture, in the original sampling procedure we perform `llm_head` and `gather` first, then perform `argmax` to full `logits`. However, we can in fact move `gather` after `argmax` to reduce both the communication volume of `gather` and the computation load of `argmax`.
<img width="1756" height="1168" alt="Image" src="https://github.com/user-attachments/assets/6dc7aeb1-3d1f-45e2-98f8-5f2ef9e35cb1" />
The test results during the puncturing phase show that this feature can optimize the `logits_processor + sampler` time consumption by more than 200 us in certain scenarios. In speculative decoding scenarios, where multiple rounds of post-processing are required for each step, the benefits of this feature can become even more pronounced. So I think this is an important optimization especially when eagle3 become more and more popular. Later I will propose a PR to implement this.
### Proposed Change.
1. Remove the `gather/all_gather` operation from logits processor.
2. Add two `gather/all_gather` operations to sampler to gather both max value and max index of `argmax`. Then perform `max` to obtain the global max value and related max index.
### Feedback Period.
_No response_
### CC List.
@youkaichao @zhuohan123 @WoosukKwon
### Any Other Things.
_No response_
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/31216 | open | [
"RFC"
] | 2025-12-23T10:23:34Z | 2025-12-26T03:33:04Z | 2 | whx-sjtu |
huggingface/diffusers | 12,884 | Compatibility issues regarding checkpoint/VAE dependency conflicts when Diffusers load Civitai LoRA | Hello everyone, I'm currently learning to use diffusers and would like to ask all my friends a question. I saw a good lora on Civitai, but this lora has requirements for checkpoint and vea. So I downloaded both models as the author requested. However, when I ran the following code, an error occurred.
The specific code is as follows:
~~~python
pipeline = StableDiffusionPipeline.from_single_file(
r"E:\Project_draw\Models\vae\clearvaeSD15_v23.safetensors",
use_safetensors=True,
torch_dtype=torch.float16,
safety_checker=None
)
~~~
The errors are as follows:
<img width="1272" height="575" alt="Image" src="https://github.com/user-attachments/assets/9788f5b9-59d2-454d-ade8-b6837115f0de" />
I checked the documentation of diffusers. The documentation mentioned that it is possible to load the model in this way, but I don't know why an error occurred. I saw many convert scripts in the script folder of diffusers, but I don't know which one is the corresponding conversion script and what the requirements are. If there are any friends who know how to solve it, could you please tell me what it feels like | https://github.com/huggingface/diffusers/issues/12884 | closed | [] | 2025-12-23T10:11:27Z | 2025-12-23T13:41:47Z | 1 | hhhFuture |
vllm-project/vllm | 31,211 | [Doc]: Add missing GPT-OSS tool calling instructions | ### 📚 The doc issue
Currently the `openai` tool calling format is not documented in [the tool calling documentation](https://docs.vllm.ai/en/stable/features/tool_calling/). However it is documented in the [cookbook](https://docs.vllm.ai/projects/recipes/en/latest/OpenAI/GPT-OSS.html#tool-use)
### Suggest a potential alternative/fix
It would make sense to list the `openai` format alongside the other formats
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/31211 | closed | [
"documentation"
] | 2025-12-23T08:35:09Z | 2025-12-25T05:29:11Z | 0 | amithkk |
huggingface/lerobot | 2,704 | Training XVLA: IndexError with auto mode; size mismatch with joint mode on 14D joint-action dataset | ### Ticket Type
🐛 Bug Report (Something isn't working)
### Environment & System Info
```Shell
```
### Description
I am trying to train XVLA with base and folding checkpoint on a 14D joint-action dataset.
When I set --policy.action_mode=auto
lerobot-train \
--dataset.repo_id= \
--output_dir=./outputs/xvla_bimanual \
--job_name=xvla_training \
--policy.dtype=bfloat16 \
--steps=3000 \
--policy.device=cuda \
--policy.action_mode=auto \
--policy.max_action_dim=20 \
--policy.repo_id= \
--policy.path="lerobot/xvla-base" \
--policy.freeze_vision_encoder=false \
--policy.freeze_language_encoder=false \
--policy.train_policy_transformer=true \
--policy.train_soft_prompts=true \
--rename_map='{
"observation.images.top": "observation.images.image",
"observation.images.right": "observation.images.image2",
"observation.images.left": "observation.images.empty_camera_0"
}'
I've got this error:
NFO 2025-12-23 06:44:14 ot_train.py:310 Output dir: outputs/xvla_bimanual
INFO 2025-12-23 06:44:14 ot_train.py:317 cfg.steps=3000 (3K)
INFO 2025-12-23 06:44:14 ot_train.py:318 dataset.num_frames=1724070 (2M)
INFO 2025-12-23 06:44:14 ot_train.py:319 dataset.num_episodes=1613
INFO 2025-12-23 06:44:14 ot_train.py:322 Effective batch size: 8 x 1 = 8
INFO 2025-12-23 06:44:14 ot_train.py:323 num_learnable_params=879482456 (879M)
INFO 2025-12-23 06:44:14 ot_train.py:324 num_total_params=879482456 (879M)
INFO 2025-12-23 06:44:14 ot_train.py:380 Start offline training on a fixed dataset, with effective batch size: 8
Traceback (most recent call last):
File "/lambda/nfs/XVLA/X-VLA/.venv/bin/lerobot-train", line 8, in <module>
sys.exit(main())
File "/lambda/nfs/XVLA/lerobot/src/lerobot/scripts/lerobot_train.py", line 516, in main
train()
File "/lambda/nfs/XVLA/lerobot/src/lerobot/configs/parser.py", line 233, in wrapper_inner
response = fn(cfg, *args, **kwargs)
File "/lambda/nfs/XVLA/lerobot/src/lerobot/scripts/lerobot_train.py", line 386, in train
batch = next(dl_iter)
File "/lambda/nfs/XVLA/lerobot/src/lerobot/datasets/utils.py", line 912, in cycle
yield next(iterator)
File "/lambda/nfs/XVLA/X-VLA/.venv/lib/python3.10/site-packages/accelerate/data_loader.py", line 579, in __iter__
next_batch = next(dataloader_iter)
File "/lambda/nfs/XVLA/X-VLA/.venv/lib/python3.10/site-packages/torch/utils/data/dataloader.py", line 733, in __next__
data = self._next_data()
File "/lambda/nfs/XVLA/X-VLA/.venv/lib/python3.10/site-packages/torch/utils/data/dataloader.py", line 1515, in _next_data
return self._process_data(data, worker_id)
File "/lambda/nfs/XVLA/X-VLA/.venv/lib/python3.10/site-packages/torch/utils/data/dataloader.py", line 1550, in _process_data
data.reraise()
File "/lambda/nfs/XVLA/X-VLA/.venv/lib/python3.10/site-packages/torch/_utils.py", line 750, in reraise
raise exception
IndexError: Caught IndexError in DataLoader worker process 2.
Original Traceback (most recent call last):
File "/lambda/nfs/XVLA/X-VLA/.venv/lib/python3.10/site-packages/torch/utils/data/_utils/worker.py", line 349, in _worker_loop
data = fetcher.fetch(index) # type: ignore[possibly-undefined]
File "/lambda/nfs/XVLA/X-VLA/.venv/lib/python3.10/site-packages/torch/utils/data/_utils/fetch.py", line 52, in fetch
data = [self.dataset[idx] for idx in possibly_batched_index]
File "/lambda/nfs/XVLA/X-VLA/.venv/lib/python3.10/site-packages/torch/utils/data/_utils/fetch.py", line 52, in <listcomp>
data = [self.dataset[idx] for idx in possibly_batched_index]
File "/lambda/nfs/XVLA/lerobot/src/lerobot/datasets/lerobot_dataset.py", line 1028, in __getitem__
item = self.hf_dataset[idx]
File "/lambda/nfs/XVLA/X-VLA/.venv/lib/python3.10/site-packages/datasets/arrow_dataset.py", line 2862, in __getitem__
return self._getitem(key)
File "/lambda/nfs/XVLA/X-VLA/.venv/lib/python3.10/site-packages/datasets/arrow_dataset.py", line 2843, in _getitem
pa_subtable = query_table(self._data, key, indices=self._indices)
File "/lambda/nfs/XVLA/X-VLA/.venv/lib/python3.10/site-packages/datasets/formatting/formatting.py", line 612, in query_table
_check_valid_index_key(key, size)
File "/lambda/nfs/XVLA/X-VLA/.venv/lib/python3.10/site-packages/datasets/formatting/formatting.py", line 552, in _check_valid_index_key
raise IndexError(f"Invalid key: {key} is out of bounds for size {size}")
IndexError: Invalid key: 1723717 is out of bounds for size 1708592
When --policy.action_mode=joint, remove --policy.max_action_dim=20, I've this error:
INFO 2025-12-23 07:14:31 ot_train.py:195 Logs will be saved locally.
INFO 2025-12-23 07:14:31 ot_train.py:207 Creating dataset
INFO 2025-12-23 07:14:32 ot_train.py:226 Creating policy
Florence2ForConditionalGeneration has generative capabilities, as `prepare_inputs_for_generation` is explicitly defined. However, it doesn't directly inherit from `GenerationMixin`. From 👉v4.50👈 onwards, `PreTrainedModel` will NOT inherit from `Gen | https://github.com/huggingface/lerobot/issues/2704 | closed | [
"bug",
"documentation",
"question",
"policies",
"dataset",
"CI",
"examples",
"training"
] | 2025-12-23T07:20:25Z | 2025-12-23T08:54:21Z | null | DaKhanh |
vllm-project/vllm | 31,205 | ValueError: Qwen3OmniMoeThinkerForConditionalGeneration does not support LoRA yet. |
hi, I have trained qwen3-omni thinker via ms-swift. However, when I tried to infer qwen3-omni with lora ckpt, an error occurred:
```
ValueError: Qwen3OmniMoeThinkerForConditionalGeneration does not support LoRA yet.
```
I have tried many verions of vllm including 0.9.2, 0.11.0 and 0.12.0
here is my script:
```
CUDA_VISIBLE_DEVICES=0,1 \
MAX_PIXELS=1003520 \
swift infer \
--model models/omni/Qwen3-Omni/Qwen3-Omni-30B-A3B-Instruct \
--adapters ckpt/Qwen3-Omni/v4-20251212-163234/checkpoint-3 \
--merge_lora false \
--stream true \
--infer_backend vllm \
--val_dataset ms-swift/data/train_test.jsonl \
--vllm_gpu_memory_utilization 0.9 \
--vllm_tensor_parallel_size 2 \
--vllm_max_model_len 32768 \
--max_new_tokens 2048 \
--vllm_limit_mm_per_prompt '{'image': 3, 'video': 3, 'audio': 3}'
```
how can I solved this problem?
| https://github.com/vllm-project/vllm/issues/31205 | open | [
"usage"
] | 2025-12-23T06:52:11Z | 2025-12-29T14:50:37Z | 2 | VJJJJJJ1 |
vllm-project/vllm | 31,204 | [RFC]: Supporting Multi MTP layers in Speculative Decoding (EagleProposer) | ### Motivation.
The EagleProposer for speculative decoding is only able to utilize the first MTP layer.
However, the model [XiaomiMiMo/MiMo-V2-Flash](https://huggingface.co/XiaomiMiMo/MiMo-V2-Flash) has 3 MTP layers.
Is there any plan or ongoing PR to extend support for multi MTP layers in speculative decoding?
btw, [hugo-wind-ding/qwq-32b-mtp](https://huggingface.co/hugo-wind-ding/qwq-32b-mtp) has 7 mtp layers for QwQ-32B
### Proposed Change.
EagleProposer needs a new member function to pass spec_step_idx to mtp models, when num_nextn_predict_layers > 1 and num_speculative_tokens > 1.
### Feedback Period.
_No response_
### CC List.
_No response_
### Any Other Things.
_No response_
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/31204 | open | [
"RFC"
] | 2025-12-23T03:34:05Z | 2025-12-23T03:34:05Z | 0 | DingYibin |
huggingface/lerobot | 2,701 | Image keys with underscores not supported when migrating to v0.4.x | ### Ticket Type
🐛 Bug Report (Something isn't working)
### Environment & System Info
```Shell
Python 3.12.3, LeRobot versions 0.3.4 and 0.4.2
From v0.4.2:
lerobot version: 0.4.2
- Platform: Linux-6.14.0-37-generic-x86_64-with-glibc2.39
- Python version: 3.12.3
- Huggingface Hub version: 0.35.3
- Datasets version: 4.1.1
- Numpy version: 2.2.6
- PyTorch version: 2.7.1+cu126
```
### Description
When upgrading a model from 0.3.4 to 0.4.2, `migrate_policy_normalization` replaces all `_` in features with `.` at https://github.com/huggingface/lerobot/blob/main/src/lerobot/processor/migrate_policy_normalization.py#L112 . I have a camera named `front_camera` used as `observations.images.front_camera`. After migrating my policy, the normalization processor expects `observations.images.front.camera`, causing https://github.com/huggingface/lerobot/blob/main/src/lerobot/processor/normalize_processor.py#L306 to fail, and my images are left unnormalized.
I've hacked it by inserting `key = key.replace("_cam",".cam")` right above the check, but this is not a good long-term fix.
### Context & Reproduction
1. Have a model trained under LeRobot 0.3.4 with non-identity normalizations on images, and the image keys having underscores, such as `front_camera`
2. Check the `input_features` in `config.json`, see the underscores in the image name
3. Migrate the model to LeRobot 0.4.2 using `migrate_policy_normalization.py`
4. See that the underscores are preserved in the migrated `config.json`, but in the corresponding `policy_preprocessor_step_<X>_normalizer_processor.safetensors` they have been replaced with dots
### Relevant logs or stack trace
```Shell
```
### Checklist
- [x] I have searched existing tickets to ensure this isn't a duplicate.
- [ ] I am using the latest version of the `main` branch.
- [x] I have verified this is not an environment-specific problem.
### Additional Info / Workarounds
I've inserted `key = key.replace("_cam",".cam")` into `normalize_processor.py` in order to make the keys match what the processor is expecting. | https://github.com/huggingface/lerobot/issues/2701 | open | [
"bug",
"question",
"policies",
"sensors",
"processor"
] | 2025-12-23T03:27:41Z | 2025-12-23T03:27:50Z | null | dangr |
huggingface/lerobot | 2,700 | Training an Smolvla model on the lerobot/aloha_sim_insertion_human dataset does not converge | ### Ticket Type
❓ Technical Question
### Environment & System Info
```Shell
Ubuntu 22.04
lerobot 0.4.1
python 3.10
lerobot-train \
--job_name aloha_smolvla \
--output_dir $OUTPUT_DIR \
--env.type=aloha \
--env.task="AlohaInsertion-v0" \
--policy.type=smolvla \
--policy.load_vlm_weights=true \
--steps=200000 \
--eval_freq=50000 \
--save_freq=50000 \
--dataset.repo_id="lerobot/aloha_sim_insertion_human" \
--policy.push_to_hub=false \
--wandb.enable=true
```
### Description
I am tring to train Smolvla model on the lerobot/aloha_sim_insertion_human dataset, but the training does not converge. In the simulation, the robotic arm trembled and got stuck at a certain position, failing to successfully pick up the object.
### Context & Reproduction
```bash
lerobot-train \
--job_name aloha_smolvla \
--output_dir $OUTPUT_DIR \
--env.type=aloha \
--env.task="AlohaInsertion-v0" \
--policy.type=smolvla \
--policy.load_vlm_weights=true \
--steps=200000 \
--eval_freq=50000 \
--save_freq=50000 \
--dataset.repo_id="lerobot/aloha_sim_insertion_human" \
--policy.push_to_hub=false \
--wandb.enable=true
lerobot-eval \
--policy.path="$CHECKPOINT_DIR" \
--env.type=aloha \
--env.task="AlohaInsertion-v0" \
--eval.n_episodes=50 \
--eval.batch_size=50
```
https://github.com/user-attachments/assets/97feea83-2f1e-45a3-9253-6dffdd13f7ea
### Relevant logs or stack trace
```Shell
```
### Checklist
- [x] I have searched existing tickets to ensure this isn't a duplicate.
- [ ] I am using the latest version of the `main` branch.
- [ ] I have verified this is not an environment-specific problem.
### Additional Info / Workarounds
_No response_ | https://github.com/huggingface/lerobot/issues/2700 | open | [
"question",
"policies",
"dataset",
"simulation",
"robots",
"training"
] | 2025-12-23T03:13:47Z | 2025-12-30T21:05:50Z | null | sslndora0612-max |
vllm-project/vllm | 31,202 | [Bug]: Mixtral Fp8 Accuracy is Degraded | ### Your current environment
H200
### 🐛 Describe the bug
- launch
```bash
vllm serve amd/Mixtral-8x7B-Instruct-v0.1-FP8-KV --enforce-eager -tp 2
```
- eval
```bash
lm_eval \
--model local-completions \
--tasks gsm8k \
--model_args "model=amd/Mixtral-8x7B-Instruct-v0.1-FP8-KV,base_url=http://localhost:8000/v1/completions,num_concurrent=1000,tokenized_requests=False"
```
- on main:
```bash
local-completions (model=amd/Mixtral-8x7B-Instruct-v0.1-FP8-KV,base_url=http://localhost:8000/v1/completions,num_concurrent=1000,tokenized_requests=False), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: 1
|Tasks|Version| Filter |n-shot| Metric | |Value | |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k| 3|flexible-extract| 5|exact_match|↑ |0.2843|± |0.0124|
| | |strict-match | 5|exact_match|↑ |0.2108|± |0.0112|
```
- on 0.12.0:
```bash
local-completions (model=amd/Mixtral-8x7B-Instruct-v0.1-FP8-KV,base_url=http://localhost:8000/v1/completions,num_concurrent=1000,tokenized_requests=False), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: 1
|Tasks|Version| Filter |n-shot| Metric | |Value | |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k| 3|flexible-extract| 5|exact_match|↑ |0.6459|± |0.0132|
| | |strict-match | 5|exact_match|↑ |0.6452|± |0.0132|
```
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/31202 | closed | [
"bug",
"help wanted"
] | 2025-12-23T02:27:28Z | 2025-12-23T02:42:58Z | 1 | robertgshaw2-redhat |
vllm-project/vllm | 31,200 | [Bug]: class Request and block_hasher has cirular reference, may cause memory leak. | ### Your current environment
<summary> Running MultiModal Network with prefix caching will cause memory leak. </summary>
<details>
<code>
class Request:
def __init__(
...
self.block_hashes: list[BlockHash] = []
self.get_hash_new_full_blocks: Callable[[], list[BlockHash]] | None = None
if block_hasher is not None:
self.get_hash_new_full_blocks = partial(block_hasher, self) # Request hold block_hasher and block_hasher hold Request, create a circular references.
self.block_hashes = self.get_hash_new_full_blocks()
</code>
Can it change to the below code?
<code>
import weakref
class Request:
def __init__(
...
self.block_hashes: list[BlockHash] = []
self.get_hash_new_full_blocks: Callable[[], list[BlockHash]] | None = None
if block_hasher is not None:
self.get_hash_new_full_blocks = partial(block_hasher, weakref.proxy(self)) # Use weakref to avoid circual references.
self.block_hashes = self.get_hash_new_full_blocks()
</code>
```text
Your output of `python collect_env.py` here
```
</details>
### 🐛 Describe the bug
Runing a multimodal, will cause memory leak.
Because the following leak trace: block_hasher -> MultiModalFeatureSpec -> MultiModalKwargsItem -> MultiModalFiledElem -> image_pixels(Tensor)
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/31200 | open | [
"bug"
] | 2025-12-23T01:55:47Z | 2025-12-23T15:02:37Z | 1 | frelam |
huggingface/diffusers | 12,881 | Is that a bug of prompt2prompt pipeline with replace word pormpt? | ### Describe the bug
It performance the same when return different cross attention map, is implement error or just the problem with prompt2prompt.
### Reproduction
Use stable-diffusion-2-1:
`images = pipe(["A turtle playing with a ball", "A monkey playing with a ball"],
generator=torch.Generator("cuda").manual_seed(34),
cross_attention_kwargs={
"edit_type": "replace",
"local_blend_words": ["turtle", "monkey"],
"n_cross_replace": 0.4,
"n_self_replace": 0.4
}).images`
It performance the same when return different cross attention map:
`class AttentionReplace(AttentionControlEdit):
def replace_cross_attention(self, attn_base, att_replace):
return attn_base.unsqueeze(0).expand(att_replace.shape[0], *attn_base.shape)
return torch.einsum("hpw,bwn->bhpn", attn_base, self.mapper)`
### Logs
```shell
```
### System Info
Diffusers=0.30.0
### Who can help?
_No response_ | https://github.com/huggingface/diffusers/issues/12881 | open | [
"bug"
] | 2025-12-23T01:55:06Z | 2025-12-23T01:55:06Z | 0 | lincion |
sgl-project/sglang | 15,641 | [Feature] In the event_loop_overlap function of the scheduler, can the recv operation be processed asynchronously? | ### Checklist
- [x] If this is not a feature request but a general question, please start a discussion at https://github.com/sgl-project/sglang/discussions. Otherwise, it will be closed.
- [x] Please use English. Otherwise, it will be closed.
### Motivation
In the _offline large-scale high-concurrency multimodal deterministic inference scenario_, when using `event_loop_overlap `on a single machine, the `recv `operation is performed **synchronously before each step**. This can cause the GPU to idle due to waiting for recv requests, **thereby reducing the utilization rate**.
We have implemented a version that moves the recv to a background thread for continuous request reception. We use a priority lock to ensure that incoming requests are prioritized for queueing and processing. _On a single machine with eight H100 cards_, we have achieved an **increase in utilization from 60% to 75%**.
We hope that sglang can provide high-quality support for large-scale offline high-concurrency scenarios with high power and high utilization requirements.
### Related resources
_No response_ | https://github.com/sgl-project/sglang/issues/15641 | open | [] | 2025-12-22T14:04:10Z | 2025-12-22T14:04:10Z | 0 | titanium-temu |
sgl-project/sglang | 15,634 | [Bug] sgl-kernel does not support fa3??? | ### Checklist
- [ ] I searched related issues but found no solution.
- [x] The bug persists in the latest version.
- [x] Issues without environment info and a minimal reproducible demo are hard to resolve and may receive no feedback.
- [x] If this is not a bug report but a general question, please start a discussion at https://github.com/sgl-project/sglang/discussions. Otherwise, it will be closed.
- [x] Please use English. Otherwise, it will be closed.
### Describe the bug
CUDA error (/sgl-kernel/build/_deps/repo-flash-attention-src/hopper/flash_fwd_launch_template.h:166): invalid configuration argument
### Reproduction
no_proxy="*" SGLANG_TORCH_PROFILER_DIR=./ python -m sglang.launch_server --model-path /root/temp_can/DeepSeek-V3-0324 --load-format dummy --tp 4 --ep 4 --disable-cuda-graph --disable-radix-cache --moe-a2a-backend deepep --deepep-mode normal --chunked-prefill-size -1 --host 0.0.0.0 --port 30000 --enable-two-batch-overlap --attention-backend fa3
### Environment
(new_py310) root@zyhuang0-0:~/temp_can/sglang# python3 -m sglang.check_env
Python: 3.10.19 (main, Oct 21 2025, 16:43:05) [GCC 11.2.0]
CUDA available: True
GPU 0,1,2,3: NVIDIA H100 80GB HBM3
GPU 0,1,2,3 Compute Capability: 9.0
CUDA_HOME: /usr/local/cuda
NVCC: Cuda compilation tools, release 12.9, V12.9.41
CUDA Driver Version: 550.54.15
PyTorch: 2.9.1+cu128
sglang: 0.5.6.post2
sgl_kernel: 0.3.19
flashinfer_python: 0.5.3
flashinfer_cubin: 0.5.3
flashinfer_jit_cache: Module Not Found
triton: 3.5.1
transformers: 4.57.1
torchao: 0.9.0
numpy: 2.2.6
aiohttp: 3.13.2
fastapi: 0.127.0
hf_transfer: 0.1.9
huggingface_hub: 0.36.0
interegular: 0.3.3
modelscope: 1.33.0
orjson: 3.11.5
outlines: 0.1.11
packaging: 25.0
psutil: 7.1.3
pydantic: 2.12.5
python-multipart: 0.0.21
pyzmq: 27.1.0
uvicorn: 0.40.0
uvloop: 0.22.1
vllm: Module Not Found
xgrammar: 0.1.27
openai: 2.6.1
tiktoken: 0.12.0
anthropic: 0.75.0
litellm: Module Not Found
decord2: 3.0.0
NVIDIA Topology:
GPU0 GPU1 GPU2 GPU3 NIC0 NIC1 NIC2 NIC3 NIC4 NIC5 NIC6 CPU Affinity NUMA Affinity GPU NUMA ID
GPU0 X NV18 NV18 NV18 SYS PIX SYS SYS SYS SYS SYS 0-47,96-143 0 N/A
GPU1 NV18 X NV18 NV18 SYS SYS SYS SYS SYS PIX SYS 48-95,144-191 1 N/A
GPU2 NV18 NV18 X NV18 SYS SYS SYS SYS SYS SYS SYS 48-95,144-191 1 N/A
GPU3 NV18 NV18 NV18 X SYS SYS SYS SYS SYS SYS PIX 48-95,144-191 1 N/A
NIC0 SYS SYS SYS SYS X SYS SYS SYS SYS SYS SYS
NIC1 PIX SYS SYS SYS SYS X SYS SYS SYS SYS SYS
NIC2 SYS SYS SYS SYS SYS SYS X PXB PXB SYS SYS
NIC3 SYS SYS SYS SYS SYS SYS PXB X PIX SYS SYS
NIC4 SYS SYS SYS SYS SYS SYS PXB PIX X SYS SYS
NIC5 SYS PIX SYS SYS SYS SYS SYS SYS SYS X SYS
NIC6 SYS SYS SYS PIX SYS SYS SYS SYS SYS SYS X
Legend:
X = Self
SYS = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
PHB = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
PXB = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
PIX = Connection traversing at most a single PCIe bridge
NV# = Connection traversing a bonded set of # NVLinks
NIC Legend:
NIC0: mlx5_0
NIC1: mlx5_1
NIC2: mlx5_2
NIC3: mlx5_3
NIC4: mlx5_4
NIC5: mlx5_5
NIC6: mlx5_6
ulimit soft: 1048576 | https://github.com/sgl-project/sglang/issues/15634 | open | [] | 2025-12-22T10:50:36Z | 2025-12-22T10:50:55Z | 0 | ziyuhuang123 |
huggingface/lerobot | 2,697 | Run pi0.5 on Libero, incorrect version of transformers | ### Ticket Type
🐛 Bug Report (Something isn't working)
### Environment & System Info
```Shell
Copy-and-paste the text below in your GitHub issue and FILL OUT the last point.
- lerobot version: 0.4.0
- Platform: Linux-6.8.0-87-generic-x86_64-with-glibc2.35
- Python version: 3.10.19
- Huggingface Hub version: 0.35.3
- Datasets version: 4.1.1
- Numpy version: 2.2.6
- PyTorch version: 2.7.1+cu126
- Is PyTorch built with CUDA support?: True
- Cuda version: 12.6
- GPU model: NVIDIA GeForce RTX 4090
- Using GPU in script?: <fill in>
```
### Description
I am running PI0.5 on Libero Benchmark and I encounter with the following issues:
```python
Built vec env | suite=libero_spatial | task_id=6 | n_envs=1
Built vec env | suite=libero_spatial | task_id=7 | n_envs=1
Built vec env | suite=libero_spatial | task_id=8 | n_envs=1
Built vec env | suite=libero_spatial | task_id=9 | n_envs=1
INFO 2025-12-22 03:42:33 bot_eval.py:499 Making policy.
The PI05 model is a direct port of the OpenPI implementation.
This implementation follows the original OpenPI structure for compatibility.
Original implementation: https://github.com/Physical-Intelligence/openpi
`torch_dtype` is deprecated! Use `dtype` instead!
Traceback (most recent call last):
File "/home/yu/miniconda3/envs/lerobot/bin/lerobot-eval", line 7, in <module>
sys.exit(main())
File "/home/yu/copy/vla/lerobot/src/lerobot/scripts/lerobot_eval.py", line 763, in main
eval_main()
File "/home/yu/copy/vla/lerobot/src/lerobot/configs/parser.py", line 233, in wrapper_inner
response = fn(cfg, *args, **kwargs)
File "/home/yu/copy/vla/lerobot/src/lerobot/scripts/lerobot_eval.py", line 501, in eval_main
policy = make_policy(
File "/home/yu/copy/vla/lerobot/src/lerobot/policies/factory.py", line 412, in make_policy
policy = policy_cls.from_pretrained(**kwargs)
File "/home/yu/copy/vla/lerobot/src/lerobot/policies/pi05/modeling_pi05.py", line 893, in from_pretrained
model = cls(config, **kwargs)
File "/home/yu/copy/vla/lerobot/src/lerobot/policies/pi05/modeling_pi05.py", line 842, in __init__
self.model = PI05Pytorch(config)
File "/home/yu/copy/vla/lerobot/src/lerobot/policies/pi05/modeling_pi05.py", line 541, in __init__
raise ValueError(msg) from None
ValueError: An incorrect transformer version is used, please create an issue on https://github.com/huggingface/lerobot/issues
```
Can I kindly ask is there anyone encounter with similar issues?
| https://github.com/huggingface/lerobot/issues/2697 | open | [
"bug",
"question",
"evaluation"
] | 2025-12-22T08:54:56Z | 2025-12-22T16:20:01Z | null | yqi19 |
huggingface/lerobot | 2,696 | RTC does not work. | ### Ticket Type
🐛 Bug Report (Something isn't working)
### Environment & System Info
```Shell
- lerobot version: 0.4.3
- Platform: Linux-5.10.134-17.3.al8.x86_64-x86_64-with-glibc2.35
- Python version: 3.10.19
- Huggingface Hub version: 0.35.3
- Datasets version: 4.1.1
- Numpy version: 2.2.6
- PyTorch version: 2.7.1+cu126
- Is PyTorch built with CUDA support?: True
- Cuda version: 12.6
- GPU model: NVIDIA H20
- Using GPU in script?: <fill in>
```
### Description
I trained pi05 using my own dataset (where the actions are absolute joint angles). The final training loss reached 0.006.
Then I ran examples/rtc/eval_dataset.py.
`rtc=RTCConfig(enabled=True, prefix_attention_schedule=<RTCAttentionSchedule.EXP: 'EXP'>, max_guidance_weight=100.0, execution_horizon=8, debug=True, debug_maxlen=1000), device='cuda:0', output_dir='rtc_debug_output', seed=42, inference_delay=4, use_torch_compile=False, torch_compile_backend='inductor', torch_compile_mode='default', torch_compile_disable_cudagraphs=True)`
<img width="2397" height="1769" alt="Image" src="https://github.com/user-attachments/assets/284b28e6-7098-4614-938b-b15582a573ce" />
<img width="3260" height="1769" alt="Image" src="https://github.com/user-attachments/assets/c1f8794a-ffaa-4ded-b4ce-04195e4011ab" />
<img width="3314" height="1769" alt="Image" src="https://github.com/user-attachments/assets/1486bbff-777a-4405-90a3-73593cd9e02b" />
<img width="3261" height="1769" alt="Image" src="https://github.com/user-attachments/assets/c9f61848-3504-4908-be94-fa8005e75039" />
<img width="3319" height="1769" alt="Image" src="https://github.com/user-attachments/assets/4dfe6dda-1108-43f4-9cd5-0d294aeaff3b" />
However, when I run the same script with the following parameters:
`
python examples/rtc/eval_dataset.py \
--policy.path=lerobot/pi05_libero_finetuned \
--dataset.repo_id=HuggingFaceVLA/libero \
--rtc.execution_horizon=8 \
--device=cuda
`
everything works fine. I would like to know where the error might be occurring.
### Context & Reproduction
The training parameters are as follows:
` "args": [
"--dataset.repo_id=/mnt/model/wlz/real_stack_purple_toy_joint_lerobot",
"--policy.type=pi05",
"--output_dir=./outputs/pi05_training",
"--job_name=pi05_training",
"--policy.repo_id=wlz",
"--policy.pretrained_path=lerobot/pi05_base",
"--policy.compile_model=true",
"--policy.gradient_checkpointing=true",
"--wandb.enable=false",
"--policy.dtype=bfloat16",
"--steps=5000",
"--policy.device=cuda",
"--batch_size=32",
"--policy.input_features={\"observation.images.image\":{\"shape\":[3,256,256],\"type\":\"VISUAL\"},\"observation.images.image2\":{\"shape\":[3,256,256],\"type\":\"VISUAL\"},\"observation.state\":{\"shape\":[7],\"type\":\"STATE\"}}",
"--policy.output_features={\"action\":{\"shape\":[7],\"type\":\"ACTION\"}}"
]
`
### Relevant logs or stack trace
```Shell
```
### Checklist
- [ ] I have searched existing tickets to ensure this isn't a duplicate.
- [ ] I am using the latest version of the `main` branch.
- [ ] I have verified this is not an environment-specific problem.
### Additional Info / Workarounds
_No response_ | https://github.com/huggingface/lerobot/issues/2696 | closed | [
"bug",
"question",
"policies",
"dataset",
"CI",
"python",
"examples",
"training"
] | 2025-12-22T03:22:23Z | 2025-12-22T05:20:39Z | null | xiaozhisky1 |
huggingface/sentence-transformers | 3,601 | how to finetuning a bi-encoder embedding model of multimodel input | I want to cluster ecommerce products by bi-encoder. For each product, it has a name(text) and an image. Can I use sentence-transfomer to finetune a bi-encoder model? The training dataset contains product clusters, like:
```
product1_name, product1_img, cluster_id1
product2_name, product2_img, cluster_id1
product3_name, product3_img, cluster_id2
productm_name,productm_img, cluster_idn
```
I want to try first to define it as a classification problem(cluster_id1,...cluster_idn) and use arcface loss. But If there are other suitable losses, it's also fine.
Is sentence transformer suitable for my use case? I find siglip(something like clip) is good at embedding. Its training data is image/text pair, but my data is not the same as it.
| https://github.com/huggingface/sentence-transformers/issues/3601 | open | [] | 2025-12-22T02:46:43Z | 2025-12-22T09:09:31Z | null | fancyerii |
vllm-project/vllm | 31,096 | [Usage]: Qwen3-Next: Both Instruct and Thinking models don't support function calling |
Does the Qwen3-Next model not support the function calling feature? Test results show some common error scenarios:
1. The tools should be called, but content returned something like the following:
```
{
"choices": [
{
"message": {
"content": "</think>\n{\"name\": \"send_email\", \"arguments\": {\"userInput\": \"ALAN的ID是123456，ALAN，中国人，英语数学很牛\"}}\n</tool_call>",
"tool_calls": []
},
"finish_reason": "stop"
}
]
}
```
```
{
"id": "chatcmpl-38af97847cce417a84577fe604d5b31e",
"object": "chat.completion",
"created": 1766119733,
"model": "Next",
"choices": [
{
"index": 0,
"message": {
"role": "assistant",
"content": "<tool_call>\n{\"name\": \"get_weather\", \"arguments\": {\"location\": \"San Francisco\", \"unit\": \"celsius\"}}\n</tool_call>",
"refusal": null,
"annotations": null,
"audio": null,
"function_call": null,
"tool_calls": [],
"reasoning_content": null
},
"logprobs": null,
"finish_reason": "stop",
"stop_reason": null,
"token_ids": null
}
],
"service_tier": null,
"system_fingerprint": null,
"usage": {
"prompt_tokens": 619,
"total_tokens": 647,
"completion_tokens": 28,
"prompt_tokens_details": null
},
"prompt_logprobs": null,
"prompt_token_ids": null,
"kv_transfer_params": null
}
```
2. Returning to the contents of the parameter list may result in non-standard characters, causing parameter retrieval to fail.
```
{
"choices": [
{
"message": {
"content": "",
"tool_calls": [
{
"id": "chatcmpl-tool-ad840dff072841759d3ed8a26e21391f",
"type": "function",
"index": 0,
"function": {
"name": "get_weather",
"arguments": ""
}
}
]
},
"finish_reason": "tool_calls"
}
]
}
```
| https://github.com/vllm-project/vllm/issues/31096 | open | [
"usage"
] | 2025-12-21T12:02:08Z | 2025-12-23T03:02:02Z | 0 | PHOEBEMOON0802 |
huggingface/lerobot | 2,694 | The GT00T algorithm simply won't run and throws the following error. Could someone please help me fix it? | The GT00T algorithm simply won't run and throws the following error. Could someone please help me fix it?
n_model.post_layernorm.bias', 'backbone.eagle_model.vision_model.vision_model.post_layernorm.weight']
Traceback (most recent call last):
File "/home/ruijia/miniconda3/envs/lerobot/bin/lerobot-train", line 7, in <module>
sys.exit(main())
File "/home/ruijia/lerobot_code/lerobot/src/lerobot/scripts/lerobot_train.py", line 517, in main
train()
File "/home/ruijia/lerobot_code/lerobot/src/lerobot/configs/parser.py", line 233, in wrapper_inner
response = fn(cfg, *args, **kwargs)
File "/home/ruijia/lerobot_code/lerobot/src/lerobot/scripts/lerobot_train.py", line 268, in train
preprocessor, postprocessor = make_pre_post_processors(
File "/home/ruijia/lerobot_code/lerobot/src/lerobot/policies/factory.py", line 252, in make_pre_post_processors
PolicyProcessorPipeline.from_pretrained(
File "/home/ruijia/lerobot_code/lerobot/src/lerobot/processor/pipeline.py", line 567, in from_pretrained
loaded_config, base_path = cls._load_config(model_id, config_filename, hub_download_kwargs)
File "/home/ruijia/lerobot_code/lerobot/src/lerobot/processor/pipeline.py", line 638, in _load_config
cls._suggest_processor_migration(model_id, f"Config file '{config_filename}' not found")
File "/home/ruijia/lerobot_code/lerobot/src/lerobot/processor/pipeline.py", line 1212, in _suggest_processor_migration
raise ProcessorMigrationError(model_path, migration_command, original_error)
lerobot.processor.pipeline.ProcessorMigrationError: Model '/home/ruijia/llmweights/GR00T-N1.5-3B' requires migration to processor format. Run: python src/lerobot/processor/migrate_policy_normalization.py --pretrained-path /home/ruijia/llmweights/GR00T-N1.5-3B
Original error: Config file 'policy_preprocessor.json' not found
@kashif @ozten @jpizarrom @julien-c @jbcayrou | https://github.com/huggingface/lerobot/issues/2694 | open | [
"bug",
"question",
"policies",
"CI",
"python",
"processor",
"examples",
"training"
] | 2025-12-21T09:12:14Z | 2025-12-24T00:06:08Z | null | wuxiaolianggit |
huggingface/lerobot | 2,693 | Wrist Roll motor not responding | ### Ticket Type
🐛 Bug Report (Something isn't working)
### Environment & System Info
```Shell
lerobot version 0.4.0
```
### Description
I connected to the lerobot so101 bot ->setup motors->callibrated->tested teleoperation
,everything wewnt fine .But after few hours when recallibration is done in some other system the wrist roll motor of the follower arm went partially stiff and it is not responding . [Used FT SCServo Debugger]
<img width="1199" height="892" alt="Image" src="https://github.com/user-attachments/assets/9c2cfbe3-b041-4847-86ef-a677b97f801a" />
### Context & Reproduction
_No response_
### Relevant logs or stack trace
```Shell
```
### Checklist
- [x] I have searched existing tickets to ensure this isn't a duplicate.
- [x] I am using the latest version of the `main` branch.
- [x] I have verified this is not an environment-specific problem.
### Additional Info / Workarounds
_No response_ | https://github.com/huggingface/lerobot/issues/2693 | open | [
"bug",
"question",
"teleoperators"
] | 2025-12-21T09:01:51Z | 2025-12-26T10:19:17Z | null | CHIRANJEET1729DAS |
huggingface/lerobot | 2,692 | [Bug] Too many errors when Train RL in Simulation | ### Ticket Type
🐛 Bug Report (Something isn't working)
### Environment & System Info
```Shell
`
- LeRobot version: 0.4.3
- Platform: Linux-6.8.0-90-generic-x86_64-with-glibc2.35
- Python version: 3.10.19
- Huggingface Hub version: 0.35.3
- Datasets version: 4.1.1
- Numpy version: 2.2.6
- FFmpeg version: N/A
- PyTorch version: 2.7.1+cu126
- Is PyTorch built with CUDA support?: True
- Cuda version: 12.6
- GPU model: NVIDIA GeForce RTX 4090 D
- Using GPU in script?: <fill in>
- lerobot scripts: ['lerobot-calibrate', 'lerobot-dataset-viz', 'lerobot-edit-dataset', 'lerobot-eval', 'lerobot-find-cameras', 'lerobot-find-joint-limits', 'lerobot-find-port', 'lerobot-imgtransform-viz', 'lerobot-info', 'lerobot-record', 'lerobot-replay', 'lerobot-setup-motors', 'lerobot-teleoperate', 'lerobot-train']
`
```
### Description
First of all, thank you for your excellent open-source work.
I've noticed that LeRobot's code has undergone some significant refactoring recently, especially the code in the hil-serl section. Therefore, I want to re-test the hil-serl code.
I followed the official documentation step by step, but I encountered many problems and don't know how to solve them. The documents: https://huggingface.co/docs/lerobot/hilserl_sim
First, I need Recording a Dataset. Therefore, I executed the following script.
```shell
python -m lerobot.rl.gym_manipulator --config_path gym_hil/env_config.json
```
The gym_hil/env_config.json is like this:
```
{
"env": {
"name": "gym_hil",
"task": "PandaPickCubeKeyboard-v0",
"fps": 10,
"robot": null,
"teleop": null,
"processor": {
"control_mode": "gamepad",
"gripper": {
"use_gripper": true,
"gripper_penalty": -0.02,
"gripper_penalty_in_reward": false
},
"reset": {
"fixed_reset_joint_positions": [0.0, 0.195, 0.0, -2.43, 0.0, 2.62, 0.785],
"reset_time_s": 2.0,
"control_time_s": 15.0,
"terminate_on_success": true
}
}
},
"dataset": {
"repo_id": "franka_sim_pick_lift_6",
"root": "/mnt/hukongtao/codebase/lerobot/franka_sim_pick_lift_6",
"task": "PandaPickCubeKeyboard-v0",
"num_episodes_to_record": 30,
"replay_episode": 0,
"push_to_hub": false
},
"mode": "record",
"device": "cpu"
}
```
from https://huggingface.co/api/resolve-cache/datasets/lerobot/config_examples/e9cea127f440dab0eb333f8b8007828ce8f48e23/rl%2Fgym_hil%2Fenv_config.json?%2Fdatasets%2Flerobot%2Fconfig_examples%2Fresolve%2Fmain%2Frl%2Fgym_hil%2Fenv_config.json=&etag=%22a4b2ef62f6cee4e31f608639134c07c7e8d3c4ab%22
I got my first error:
```
draccus.utils.DecodingError: `processor.gripper`: Could not decode the value into any of the given types:
GripperConfig: The fields `gripper_penalty_in_reward` are not valid for GripperConfig
```
So I delete gripper_penalty_in_reward in gym_hil/env_config.json. Then I ran the program again. I got another error:
```
Traceback (most recent call last):
File "/data/hukongtao/miniconda3/envs/lerobot/lib/python3.10/runpy.py", line 196, in _run_module_as_main
return _run_code(code, main_globals, None,
File "/data/hukongtao/miniconda3/envs/lerobot/lib/python3.10/runpy.py", line 86, in _run_code
exec(code, run_globals)
File "/mnt/hukongtao/codebase/lerobot/src/lerobot/rl/gym_manipulator.py", line 770, in <module>
main()
File "/mnt/hukongtao/codebase/lerobot/src/lerobot/configs/parser.py", line 233, in wrapper_inner
response = fn(cfg, *args, **kwargs)
File "/mnt/hukongtao/codebase/lerobot/src/lerobot/rl/gym_manipulator.py", line 766, in main
control_loop(env, env_processor, action_processor, teleop_device, cfg)
File "/mnt/hukongtao/codebase/lerobot/src/lerobot/rl/gym_manipulator.py", line 602, in control_loop
action_features = teleop_device.action_features
AttributeError: 'NoneType' object has no attribute 'action_features'
```
I'm certain this is a bug in the code, because in a simulation environment teleop_device is None. But I don't know how to fix this error.
### Context & Reproduction
```
git clone https://github.com/huggingface/lerobot.git
cd lerobot
pip install -e .[all]
python -m lerobot.rl.gym_manipulator --config_path gym_hil/env_config.json
```
### Relevant logs or stack trace
```Shell
```
### Checklist
- [x] I have searched existing tickets to ensure this isn't a duplicate.
- [x] I am using the latest version of the `main` branch.
- [x] I have verified this is not an environment-specific problem.
### Additional Info / Workarounds
I tested the hil-serl code in version 0.3.3 of lerobot, and it worked without any problems. | https://github.com/huggingface/lerobot/issues/2692 | open | [
"bug",
"documentation",
"question",
"dataset",
"simulation",
"tests",
"examples",
"training"
] | 2025-12-21T08:22:16Z | 2026-01-04T06:19:05Z | null | Hukongtao |
huggingface/accelerate | 3,894 | How to specify different number of process per node | I've 2 node. First node has 8 gpus while second node has 2 GPUs. I want to specify the number of process to be 8 and 2 respectively in both nodes. I'm using this config in both node. But it always tries to divide equal number of process in both node. With below config file, it's starting 5 process in both nodes:-
Node 1:-
```compute_environment: LOCAL_MACHINE
debug: false
distributed_type: MULTI_GPU
downcast_bf16: 'no'
enable_cpu_affinity: false
gpu_ids: 0,1,2,3,4,5,6,7
machine_rank: 0
main_process_ip: xxxxx
main_process_port: 5000
main_training_function: main
mixed_precision: fp16
num_machines: 2
num_processes: 10
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false
```
Node 2:-
```
compute_environment: LOCAL_MACHINE
debug: false
distributed_type: MULTI_GPU
downcast_bf16: 'no'
enable_cpu_affinity: false
gpu_ids: 0,1
machine_rank: 1
main_process_ip: xxxx
main_process_port: 5000
main_training_function: main
mixed_precision: fp16
num_machines: 2
num_processes: 10
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false
``` | https://github.com/huggingface/accelerate/issues/3894 | open | [] | 2025-12-21T07:09:15Z | 2025-12-21T07:09:15Z | null | AIML001 |
vllm-project/vllm | 31,091 | [Usage]: Image Embedding Models (CLIP, Siglip, etc) | ### Your current environment
```text
root@3904bdeddb91:/vllm-workspace# python3 collect_env.py
Collecting environment information...
==============================
System Info
==============================
OS : Ubuntu 22.04.5 LTS (x86_64)
GCC version : (Ubuntu 11.4.0-1ubuntu1~22.04.2) 11.4.0
Clang version : Could not collect
CMake version : Could not collect
Libc version : glibc-2.35
==============================
PyTorch Info
==============================
PyTorch version : 2.9.0+cu129
Is debug build : False
CUDA used to build PyTorch : 12.9
ROCM used to build PyTorch : N/A
==============================
Python Environment
==============================
Python version : 3.12.12 (main, Oct 10 2025, 08:52:57) [GCC 11.4.0] (64-bit runtime)
Python platform : Linux-6.8.0-87-generic-x86_64-with-glibc2.35
==============================
CUDA / GPU Info
==============================
Is CUDA available : True
CUDA runtime version : 12.9.86
CUDA_MODULE_LOADING set to :
GPU models and configuration :
GPU 0: NVIDIA RTX PRO 6000 Blackwell Workstation Edition
GPU 1: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition
Nvidia driver version : 580.65.06
cuDNN version : Could not collect
HIP runtime version : N/A
MIOpen runtime version : N/A
Is XNNPACK available : True
==============================
CPU Info
==============================
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 43 bits physical, 48 bits virtual
Byte Order: Little Endian
CPU(s): 64
On-line CPU(s) list: 0-63
Vendor ID: AuthenticAMD
Model name: AMD EPYC 7502 32-Core Processor
CPU family: 23
Model: 49
Thread(s) per core: 2
Core(s) per socket: 32
Socket(s): 1
Stepping: 0
Frequency boost: enabled
CPU max MHz: 2500.0000
CPU min MHz: 1500.0000
BogoMIPS: 4999.95
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip rdpid overflow_recov succor smca sev sev_es ibpb_exit_to_user
Virtualization: AMD-V
L1d cache: 1 MiB (32 instances)
L1i cache: 1 MiB (32 instances)
L2 cache: 16 MiB (32 instances)
L3 cache: 128 MiB (8 instances)
NUMA node(s): 1
NUMA node0 CPU(s): 0-63
Vulnerability Gather data sampling: Not affected
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Not affected
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed: Mitigation; untrained return thunk; SMT enabled with STIBP protection
Vulnerability Spec rstack overflow: Mitigation; Safe RET
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Retpolines; IBPB conditional; STIBP always-on; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
Vulnerability Vmscape: Mitigation; IBPB before exit to userspace
==============================
Versions of relevant libraries
==============================
[pip3] flashinfer-python==0.5.3
| https://github.com/vllm-project/vllm/issues/31091 | closed | [
"usage"
] | 2025-12-21T04:10:10Z | 2025-12-23T03:26:40Z | 2 | JamesDConley |
huggingface/lerobot | 2,690 | [Bug] Pi0 Inference RuntimeError: Dimension mismatch in Gemma eager_attention_forward (Causal Mask vs Attn Weights) | https://github.com/huggingface/lerobot/issues/2690 | closed | [
"bug",
"question",
"policies",
"dataset",
"CI",
"performance",
"robots",
"examples",
"training"
] | 2025-12-20T16:08:36Z | 2025-12-22T09:34:57Z | null | SMWTDDY | |
huggingface/lerobot | 2,689 | problem regarding to update aloha sim dataset version v2.1 to v3.0 | ### Ticket Type
🐛 Bug Report (Something isn't working)
### Environment & System Info
```Shell
lerobot version 3.0, h100 gpu, openpi repository, training aloha simulation with pi0.5
```
### Description
During training aloha simulation, I updated lerobot aloha sim insertion dataset from compatible with 2.1 to 3.0, the training results showing aloha joints are working weirdly (showing spark of joint actions).
The dataset conversion followed as below.
```
lerobot.datasets.backward_compatibility.BackwardCompatibilityError:
The dataset you requested (lerobot/aloha_sim_insertion_scripted) is in 2.1 format.
We introduced a new format since v3.0 which is not backward compatible with v2.1.
Please, update your dataset to the new format using this command:
python -m lerobot.datasets.v30.convert_dataset_v21_to_v30 --repo-id=lerobot/aloha_sim_insertion_scripted
```
### Context & Reproduction
_No response_
### Relevant logs or stack trace
```Shell
```
### Checklist
- [ ] I have searched existing tickets to ensure this isn't a duplicate.
- [ ] I am using the latest version of the `main` branch.
- [ ] I have verified this is not an environment-specific problem.
### Additional Info / Workarounds
_No response_ | https://github.com/huggingface/lerobot/issues/2689 | open | [
"bug",
"question",
"dataset",
"simulation",
"CI",
"robots",
"training"
] | 2025-12-20T13:42:39Z | 2025-12-24T00:06:09Z | null | conscious-choi |
sgl-project/sglang | 15,524 | [Bug] Deepseek R1 multi-turn tool calling not working | ### Checklist
- [x] I searched related issues but found no solution.
- [x] The bug persists in the latest version.
- [ ] Issues without environment info and a minimal reproducible demo are hard to resolve and may receive no feedback.
- [x] If this is not a bug report but a general question, please start a discussion at https://github.com/sgl-project/sglang/discussions. Otherwise, it will be closed.
- [x] Please use English. Otherwise, it will be closed.
### Describe the bug
The multi-turn tool calling failed with error: `{"object":"error","message":"'dict object' has no attribute 'name'","type":"BadRequest","param":null,"code":400}`
Here is the example query:
```
curl http://127.0.0.1:7080/v1/chat/completions -H "Content-Type: application/json" -d '{
"model": "deepseek-ai/DeepSeek-R1",
"stream": false,
"messages": [
{
"role": "user",
"content": "What is the weather like in San Francisco?"
},
{
"role": "assistant",
"content": "I will check the weather for San Francisco. Please hold on.",
"tool_calls": [
{
"id": "call_ab97cb439a5e41cfbdd8960c",
"type": "function",
"function": {
"name": "get_weather",
"arguments": "{\"location\": \"San Francisco, CA\"}"
}
}
]
},
{
"role": "tool",
"tool_call_id": "call_ab97cb439a5e41cfbdd8960c",
"content": "70 degrees and foggy"
}
],
"tools": [
{
"type": "function",
"function": {
"name": "get_weather",
"description": "Get the current weather",
"parameters": {
"type": "object",
"properties": {
"location": {
"type": "string",
"description": "The city and state (both required), e.g. San Francisco, CA."
}
},
"required": [
"location"
]
}
}
}
]
}'
```
However, the same query works for the image back in August.
### Reproduction
*) Start server on B200
```
python3 -m sglang.launch_server \
--model-path nvidia/DeepSeek-R1-0528-NVFP4 \
--port 7080 \
--host 0.0.0.0 \
--tp-size=8 \
--ep-size=8 \
--moe-runner-backend=flashinfer_trtllm \
--enable-flashinfer-allreduce-fusion \
--tool-call-parser=deepseekv3 \
--chat-template=/sgl-workspace/sglang/examples/chat_template/tool_chat_template_deepseekr1.jinja \
--speculative-num-steps=3 \
--speculative-eagle-topk=1 \
--speculative-num-draft-tokens=4 \
--speculative-algorithm=EAGLE \
--trust-remote-code
```
*) send query
```
curl http://127.0.0.1:7080/v1/chat/completions -H "Content-Type: application/json" -d '{
"model": "deepseek-ai/DeepSeek-R1",
"stream": false,
"messages": [
{
"role": "user",
"content": "What is the weather like in San Francisco?"
},
{
"role": "assistant",
"content": "I will check the weather for San Francisco. Please hold on.",
"tool_calls": [
{
"id": "call_ab97cb439a5e41cfbdd8960c",
"type": "function",
"function": {
"name": "get_weather",
"arguments": "{\"location\": \"San Francisco, CA\"}"
}
}
]
},
{
"role": "tool",
"tool_call_id": "call_ab97cb439a5e41cfbdd8960c",
"content": "70 degrees and foggy"
}
],
"tools": [
{
"type": "function",
"function": {
"name": "get_weather",
"description": "Get the current weather",
"parameters": {
"type": "object",
"properties": {
"location": {
"type": "string",
"description": "The city and state (both required), e.g. San Francisco, CA."
}
},
"required": [
"location"
]
}
}
}
]
}'
```
### Environment
```
Python: 3.12.3 (main, Nov 6 2025, 13:44:16) [GCC 13.3.0]
CUDA available: True
GPU 0,1,2,3,4,5,6,7: NVIDIA B200
GPU 0,1,2,3,4,5,6,7 Compute Capability: 10.0
CUDA_HOME: /usr/local/cuda
NVCC: Cuda compilation tools, release 12.9, V12.9.86
CUDA Driver Version: 580.95.05
PyTorch: 2.9.1+cu129
sglang: 0.5.6.post2
sgl_kernel: 0.3.19
flashinfer_python: 0.5.3
flashinfer_cubin: 0.5.3
flashinfer_jit_cache: Module Not Found
triton: 3.5.1
transformers: 4.57.1
torchao: 0.9.0
numpy: 2.3.5
aiohttp: 3.13.2
fastapi: 0.124.2
hf_transfer: 0.1.9
huggingface_hub: 0.36.0
interegular: 0.3.3
modelscope: 1.33.0
orjson: 3.11.5
outlines: 0.1.11
packaging: 25.0
psutil: 7.1.3
pydantic: 2.12.5
python-multipart: 0.0.20
pyzmq: 27.1.0
uvicorn: 0.38.0
uvloop: 0.22.1
vllm: Module Not Found
xgrammar: 0.1.27
openai: 2.6.1
tiktoken: 0.12.0
anthropic: 0.75.0
litellm: Module Not Found
decord2: 2.0.0
NVIDIA Topology:
GPU0 GPU1 GPU2 GPU3 GPU4 GPU5 GPU6 GPU7 CPU Affinity NUMA Affinity GPU NUMA ID
GPU0 X NV18 NV18 NV18 NV18 NV18 NV18 NV18 0-55,112-167 0 N/A
GPU1 NV18 X NV18 NV18 NV18 NV18 NV18 NV1 | https://github.com/sgl-project/sglang/issues/15524 | closed | [] | 2025-12-20T10:31:36Z | 2025-12-21T01:29:43Z | 2 | ynwang007 |
vllm-project/vllm | 31,066 | [Doc]: Formatting issue in markdown file | ### 📚 The doc issue
in [paged_attention.md](https://github.com/vllm-project/vllm/blob/ff2168bca3a195b835c64a5c9012d7b6a9f34e61/docs/design/paged_attention.md#query), there is an issue where a pictures arent formatted correctly and only show the html link .
For example, specifically, in the Query subsection, we can see:
`{ align="center" alt="q_vecs" width="70%" }`
The asset isnt loaded correctly.
There are a total of **7 such issues**, particularly, we have
- Query subsection - 2 instances.
- Key subsection - 2 instances.
- Value subsection - 3 instances
### Suggest a potential alternative/fix
Perhaps the reference for the images can be checked, it must be broken somewhere
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/31066 | closed | [
"documentation"
] | 2025-12-20T06:23:44Z | 2025-12-22T01:38:56Z | 1 | ssaketh-ch |
vllm-project/vllm | 31,044 | [CI Failure]: Blackwell Fusion Tests | ### Name of failing test
FAILED tests/compile/test_fusion_attn.py::test_attention_quant_pattern[AttentionBackendEnum.TRITON_ATTN-nvidia/Llama-4-Scout-17B-16E-Instruct-FP8-TestAttentionFp8StaticQuantPatternModel--quant_fp8-dtype1-533-128-40-8] - AssertionError: Tensor-likes are not close!
### Basic information
- [x] Flaky test
- [ ] Can reproduce locally
- [ ] Caused by external libraries (e.g. bug in `transformers`)
### 🧪 Describe the failing test
On B200:
FAILED tests/compile/test_fusion_attn.py::test_attention_quant_pattern[AttentionBackendEnum.TRITON_ATTN-nvidia/Llama-4-Scout-17B-16E-Instruct-FP8-TestAttentionFp8StaticQuantPatternModel--quant_fp8-dtype1-533-128-40-8] - AssertionError: Tensor-likes are not close!
```bash
pytest -v -x tests/compile/test_fusion_attn.py::test_attention_quant_pattern
```
### 📝 History of failing test
x
### CC List.
x | https://github.com/vllm-project/vllm/issues/31044 | open | [
"help wanted",
"torch.compile",
"ci-failure"
] | 2025-12-19T18:49:59Z | 2025-12-26T21:58:25Z | 3 | robertgshaw2-redhat |
vllm-project/vllm | 31,043 | [BugFix]: move torch.Size across graphs in split_graph | ### 🚀 The feature, motivation and pitch
When fixing a moe x cudagraph issue (see #30914), we found that `split_graph` may generate a submodule that returns a torch.Size and later another submodule that takes torch.Size. This errors since pt2 somehow does not support `torch.Size` as output yet.
One fix is to manually reorder some lines in the model code to avoid this split happen between getting the `torch.Size` and using it. But this is too intrusive and requires manual efforts on many models.
A more automated approach is to have a graph pass in `split_graph` to move the torch.Size a bit to avoid patterns like
```
# Old:
size = tensor_a.shape
some_cg_unsafe_op
tensor_b = tensor_b.view(size)
```
---->
```
# New:
some_cg_unsafe_op
size = tensor_a.shape
tensor_b = tensor_b.view(size)
```
### Alternatives
_No response_
### Additional context
_No response_
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/31043 | open | [
"help wanted",
"feature request",
"torch.compile"
] | 2025-12-19T18:24:58Z | 2025-12-22T21:23:04Z | 1 | BoyuanFeng |
vllm-project/vllm | 31,039 | [Feature]: Integrate Sonic MoE | ### 🚀 The feature, motivation and pitch
https://x.com/wentaoguo7/status/2001773245318541324?s=46&t=jLcDgQXDbYe6HgFmTNYgpg
https://github.com/Dao-AILab/sonic-moe
Curious to see benchmarks!
### Alternatives
_No response_
### Additional context
_No response_
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/31039 | open | [
"help wanted",
"good first issue",
"feature request"
] | 2025-12-19T17:29:59Z | 2026-01-04T14:10:21Z | 4 | robertgshaw2-redhat |
sgl-project/sglang | 15,481 | [Bug] Seeded Deterministic/Batch Invariant Inference Not Working on v1/completions endpoint | ### Checklist
- [x] I searched related issues but found no solution.
- [x] The bug persists in the latest version.
- [x] Issues without environment info and a minimal reproducible demo are hard to resolve and may receive no feedback.
- [x] If this is not a bug report but a general question, please start a discussion at https://github.com/sgl-project/sglang/discussions. Otherwise, it will be closed.
- [x] Please use English. Otherwise, it will be closed.
### Describe the bug
I’m trying to enable batch-invariant (deterministic) inference while serving SGLang behind an OpenAI API-compatible interface.
Deterministic inference docs: https://docs.sglang.io/advanced_features/deterministic_inference.html
## What works
The native /generate endpoint correctly varies output by seed and is repeatable per seed.
Example request:
POST {base}/generate
```json
{
"text": "generate a uuid. UUID:",
"sampling_params": {
"temperature": 1,
"max_new_tokens": 32,
"sampling_seed": 0
}
}
```
Behavior: changing sampling_seed changes the output; repeating with the same sampling_seed reproduces it.
## What doesn’t work
On the OpenAI-compatible endpoint POST {base}/v1/completions, seed appears to have no effect (even with temperature=1 and top_p=1).
Example:
POST {base}/v1/completions
```json
{
"model": "Qwen/Qwen3-30B-A3B",
"prompt": "generate a uuid. UUID: ",
"max_tokens": 32,
"temperature": 1,
"top_p": 1,
"n": 1,
"seed": 0
}
```
Behavior: response is the same regardless of seed value.
Expected behavior
With --enable-deterministic-inference, I expected the OpenAI-compatible endpoints to:
* honor seed as the sampling seed (analogous to sampling_seed), and
* remain deterministic/repeatable for the same (prompt, params, seed).
### Reproduction
Server launch:
```bash
exec python3 -m sglang.launch_server \
--model-path "Qwen/Qwen3-30B-A3B" \
--host 0.0.0.0 \
--port 8000 \
--tp "1" \
--attention-backend "triton" \
--context-length "32000" \
--trust-remote-code \
--enable-deterministic-inference
```
POST {base}/v1/completions
```json
{
"model": "Qwen/Qwen3-30B-A3B",
"prompt": "generate a uuid. UUID: ",
"max_tokens": 32,
"temperature": 1,
"top_p": 1,
"n": 1,
"seed": 0
}
```
varying the seed results in same output
### Environment
==========
== CUDA ==
==========
CUDA Version 12.9.1
Container image Copyright (c) 2016-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
This container image and its contents are governed by the NVIDIA Deep Learning Container License.
By pulling and using the container, you accept the terms and conditions of this license:
https://developer.nvidia.com/ngc/nvidia-deep-learning-container-license
A copy of this license is made available in this container at /NGC-DL-CONTAINER-LICENSE for your convenience.
Auto-detected 1 GPU(s)
Python: 3.12.3 (main, Nov 6 2025, 13:44:16) [GCC 13.3.0]
CUDA available: True
GPU 0: NVIDIA RTX PRO 6000 Blackwell Server Edition
GPU 0 Compute Capability: 12.0
CUDA_HOME: /usr/local/cuda
NVCC: Cuda compilation tools, release 12.9, V12.9.86
CUDA Driver Version: 580.105.08
PyTorch: 2.9.1+cu129
sglang: 0.5.6.post2
sgl_kernel: 0.3.19
flashinfer_python: 0.5.3
flashinfer_cubin: 0.5.3
flashinfer_jit_cache: Module Not Found
triton: 3.5.1
transformers: 4.57.1
torchao: 0.9.0
numpy: 2.3.5
aiohttp: 3.13.2
fastapi: 0.124.2
hf_transfer: 0.1.9
huggingface_hub: 0.36.0
interegular: 0.3.3
modelscope: 1.33.0
orjson: 3.11.5
outlines: 0.1.11
packaging: 25.0
psutil: 7.1.3
pydantic: 2.12.5
python-multipart: 0.0.20
pyzmq: 27.1.0
uvicorn: 0.38.0
uvloop: 0.22.1
vllm: Module Not Found
xgrammar: 0.1.27
openai: 2.6.1
tiktoken: 0.12.0
anthropic: 0.75.0
litellm: Module Not Found
decord2: 2.0.0
NVIDIA Topology:
[4mGPU0 NIC0 CPU Affinity NUMA Affinity GPU NUMA ID[0m
GPU0 X SYS 0-63,128-191 0 N/A
NIC0 SYS X
Legend:
X = Self
SYS = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
PHB = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
PXB = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
PIX = Connection traversing at most a single PCIe bridge
NV# = Connection traversing a bonded set of # NVLinks
NIC Legend:
NIC0: mlx5_bond_0
ulimit soft: 1024 | https://github.com/sgl-project/sglang/issues/15481 | closed | [
"bug",
"high priority"
] | 2025-12-19T15:04:26Z | 2025-12-20T04:32:15Z | 8 | jamesheavey |
huggingface/lerobot | 2,684 | How to manually push a dataset | Say you `lerobot-record` a dataset with the flag `--dataset.push_to_hub=False`, or you encounter any problem at uploading time.
Is using `hf upload` enough, or does `lerobot` datasets need additional stuff? | https://github.com/huggingface/lerobot/issues/2684 | open | [
"documentation",
"question",
"dataset"
] | 2025-12-19T13:00:20Z | 2025-12-19T15:41:42Z | null | mcres |
vllm-project/vllm | 31,023 | [Doc]: FP8 KV Cache: Does softmax output multiply with FP8 V directly or after dequantization? | ### 📚 The doc issue
https://docs.vllm.ai/en/v0.8.5.post1/features/quantization/quantized_kvcache.html
Question:
In the FP8 KV Cache implementation, after computing attention scores and softmax at higher precision (FP16/BF16), is the resulting attention weight matrix:
Quantized to FP8 and multiplied directly with FP8 V cache, or
Multiplied with V cache after dequantizing V to higher precision?
The documentation mentions "no fused dequantization and attention operations yet" but doesn't specify the precision of this final multiplication. Clarifying this detail would help understand the accuracy-performance tradeoff.
Thanks!
### Suggest a potential alternative/fix
_No response_
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/31023 | closed | [
"documentation"
] | 2025-12-19T10:33:22Z | 2025-12-22T00:41:38Z | 0 | jorjiang |
vllm-project/vllm | 31,019 | [Bug]: Qwen3-VL 2:4 sparsity llm-compressor RuntimeError: shape mismatch (0.12, 0.13rc2) | ### Your current environment
<details>
<summary>The output of <code>python collect_env.py</code></summary>
```text
Collecting environment information...
==============================
System Info
==============================
OS : Ubuntu 24.04.3 LTS (x86_64)
GCC version : (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
Clang version : Could not collect
CMake version : Could not collect
Libc version : glibc-2.39
==============================
PyTorch Info
==============================
PyTorch version : 2.9.0+cu128
Is debug build : False
CUDA used to build PyTorch : 12.8
ROCM used to build PyTorch : N/A
==============================
Python Environment
==============================
Python version : 3.12.3 (main, Nov 6 2025, 13:44:16) [GCC 13.3.0] (64-bit runtime)
Python platform : Linux-6.14.0-1017-azure-x86_64-with-glibc2.39
==============================
CUDA / GPU Info
==============================
Is CUDA available : True
CUDA runtime version : 12.8.93
CUDA_MODULE_LOADING set to :
GPU models and configuration : GPU 0: NVIDIA H100 NVL
Nvidia driver version : 580.95.05
cuDNN version : Could not collect
HIP runtime version : N/A
MIOpen runtime version : N/A
Is XNNPACK available : True
==============================
CPU Info
==============================
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 48 bits physical, 48 bits virtual
Byte Order: Little Endian
CPU(s): 40
On-line CPU(s) list: 0-39
Vendor ID: AuthenticAMD
Model name: AMD EPYC 9V84 96-Core Processor
CPU family: 25
Model: 17
Thread(s) per core: 1
Core(s) per socket: 40
Socket(s): 1
Stepping: 1
BogoMIPS: 4800.09
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl xtopology tsc_reliable nonstop_tsc cpuid extd_apicid aperfmperf tsc_known_freq pni pclmulqdq ssse3 fma cx16 pcid sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand hypervisor lahf_lm cmp_legacy cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw topoext perfctr_core vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves user_shstk avx512_bf16 clzero xsaveerptr rdpru arat avx512vbmi umip avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq rdpid fsrm
Hypervisor vendor: Microsoft
Virtualization type: full
L1d cache: 1.3 MiB (40 instances)
L1i cache: 1.3 MiB (40 instances)
L2 cache: 40 MiB (40 instances)
L3 cache: 160 MiB (5 instances)
NUMA node(s): 1
NUMA node0 CPU(s): 0-39
Vulnerability Gather data sampling: Not affected
Vulnerability Ghostwrite: Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Not affected
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed: Not affected
Vulnerability Spec rstack overflow: Vulnerable: Safe RET, no microcode
Vulnerability Spec store bypass: Vulnerable
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Retpolines; STIBP disabled; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds: Not affected
Vulnerability Tsa: Vulnerable: Clear CPU buffers attempted, no microcode
Vulnerability Tsx async abort: Not affected
Vulnerability Vmscape: Not affected
==============================
Versions of relevant libraries
==============================
[pip3] flashinfer-python==0.5.3
[pip3] numpy==2.2.6
[pip3] nvidia-cublas-cu12==12.8.4.1
[pip3] nvidia-cuda-cupti-cu12==12.8.90
[pip3] nvidia-cuda-nvrtc-cu12==12.8.93
[pip3] nvidia-cuda-runtime | https://github.com/vllm-project/vllm/issues/31019 | open | [
"bug",
"help wanted",
"good first issue"
] | 2025-12-19T09:18:00Z | 2025-12-24T12:16:01Z | 4 | SorenDreano |
vllm-project/vllm | 31,016 | [Bug]: FlashInfer Incompatible with Sleep Mode | ### Your current environment
<details>
<summary>The output of <code>python collect_env.py</code></summary>
```text
Your output of `python collect_env.py` here
```
</details>
### 🐛 Describe the bug
Here is a script to reproduce the bug:
I use vllm=v0.10.1 and flashinfer-python=v0.5.3.
```
from vllm import LLM, SamplingParams
if __name__ == "__main__":
model_pth = "xxx/Qwen3-1.7B"
tp_size = 1
llm = LLM(
model=model_pth,
enable_sleep_mode=True,
tensor_parallel_size=tp_size,
gpu_memory_utilization=0.7,
)
llm.sleep(level=1)
llm.wake_up()
prompts = [
"What is AI?",
"Where is the Machu Picchu located?",
"What is the capital of France?",
"Who painted the Mona Lisa?",
]
sampling_params = SamplingParams(
temperature=0.7,
top_p=0.9,
max_tokens=64,
)
outputs = llm.generate(prompts, sampling_params)
for i, out in enumerate(outputs):
prompt = prompts[i]
generated = out.outputs[0].text
print(f"Prompt {i}: {prompt!r}")
print(f"Generation: {generated}\n")
```
### Root Cause
The bug occurs because the FlashInfer backend’s `attn_metadata` is stateful. It holds a `block_table_arange` tensor that is initialized once and then reused across subsequent calls to `build`:
```python
self.block_table_arange = torch.arange(
max_num_pages_per_req,
dtype=torch.int32,
device=self.device,
)
```
This `block_table_arange` tensor is allocated in the mempool with the `"kv_cache"` tag. It gets discarded after calling `llm.sleep`, but is not recreated when the engine wakes up, which leads to incorrect values and thus wrong outputs.
Specifically, this will cause bad rollout outputs in VERL using vllm + flashinfer.
### Temporary Fix
Here is a patch as a temporary workaround. It’s not an ideal solution, but it works:
```python
from vllm.v1.attention.backends.flashinfer import FlashInferMetadataBuilder
import torch
def patch_flashinfer_build():
old_build = FlashInferMetadataBuilder.build
def new_build(*args, **kwargs):
self = args[0]
max_num_pages_per_req = self.block_table_arange.numel()
self.block_table_arange.copy_(
torch.arange(
max_num_pages_per_req,
device=self.block_table_arange.device,
dtype=self.block_table_arange.dtype,
)
)
return old_build(*args, **kwargs)
FlashInferMetadataBuilder.build = new_build
patch_flashinfer_build()
```
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/31016 | open | [
"bug",
"help wanted"
] | 2025-12-19T08:04:19Z | 2025-12-19T23:17:47Z | 1 | xiaoxiaosuaxuan |
huggingface/transformers.js | 1,490 | Example models for each pipeline | ### Question
Right now, I sorta use the docs and some searches to find good default models for https://workglow.dev/ for each pipeline that transformerjs has to offer. But they are not really the best, either in size or performance.
It would be great to have a list for each pipeline for fast and effective, best of breed, and a workhorse that is in between. Like a good, better, best. | https://github.com/huggingface/transformers.js/issues/1490 | open | [
"question"
] | 2025-12-19T07:37:16Z | 2025-12-19T17:41:01Z | null | sroussey |
vllm-project/vllm | 31,004 | [New Model]: T5Gemma 2 | ### The model to consider.
https://huggingface.co/collections/google/t5gemma-2
### The closest model vllm already supports.
_No response_
### What's your difficulty of supporting the model you want?
I know vLLM dropped encoder-decoder support, but can we bring it back?
https://huggingface.co/docs/transformers/model_doc/t5gemma2
https://blog.google/technology/developers/t5gemma-2/
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/31004 | open | [
"new-model"
] | 2025-12-19T03:55:00Z | 2025-12-20T21:37:34Z | 1 | ducviet00-h2 |
sgl-project/sglang | 15,443 | SGLang Diffusion Cookbook Proposal | # 🎨 [Community Contribution] Create SGLang Diffusion Models Cookbook
## 🎯 Goal
Create a comprehensive cookbook for diffusion models in SGLang, demonstrating SGLang's performance advantages for image and video generation workloads.
## 📋 Scope
### Models to Cover
**Image Generation:**
- Flux-1 Dev
- Flux-2
- SDXL-Turbo
- Qwen Image Edit
**Video Generation:**
- Wan 2.1
- Wan 2.2
### Content Structure
Each model section includes:
1. **Model Introduction**
- Capabilities and use cases
- Resolution/quality specifications
- Style examples and output samples
- Links to official resources
2. **SGLang Deployment**
- One-command server launch
- Client usage example
- Model-specific optimization tips
3. **Performance Benchmarks**
- Throughput (images/sec or videos/min)
- Latency and memory usage
- Comparison: SGLang vs Diffusers vs ComfyUI
- Bar charts and scaling analysis
- Reproducible benchmark scripts
## 📦 Deliverables
```
cookbook/diffusion/
├── README.md # Main cookbook
├── examples/ # Usage scripts per model
│ ├── flux1_basic.py
│ ├── sdxl_turbo.py
│ ├── wan21_video.py
│ └── ...
├── benchmarks/
│ ├── bench_image.py
│ ├── bench_video.py
│ ├── compare_backends.py
│ └── run_all.sh
└── assets/
└── output_examples/ # Curated generation examples
```
## 🚀 Timeline
**Phase 1 (Weeks 1-2):** MVP with Flux-1 + SDXL-Turbo
**Phase 2 (Weeks 3-4):** Add remaining image models
**Phase 3 (Weeks 5-6):** Video models + comprehensive benchmarks
## 💪 How to Contribute
We need help with:
### Required Contributors (2-3 people)
- [ ] **Benchmark Engineer**: Run performance tests on H100/A100
- Time commitment: ~10 hours/week for 4 weeks
- Requirements: GPU access, Python proficiency
- [ ] **Documentation Writer**: Create usage examples and guides
- Time commitment: ~8 hours/week for 4 weeks
- Requirements: Technical writing, SGLang familiarity
- [ ] **Visual Designer** (optional): Curate output examples
- Time commitment: ~5 hours/week for 2 weeks
- Requirements: Eye for quality, prompt engineering
### Hardware Requirements
- H100 (80GB) - primary testing platform
- A100 (40GB) - secondary platform (optional)
- Access via cloud providers acceptable (AWS/Lambda/RunPod)
## 📝 Contribution Process
1. **Comment below** if interested (mention which role)
2. **Join discussion** on implementation details
3. **Fork repo** and work on assigned section
4. **Submit PR** following SGLang cookbook standards
5. **Iterate** based on review feedback
## 🔗 References
- [SGLang Cookbook Template](https://cookbook.sglang.io/)
- [DeepSeek-V3 Example](https://cookbook.sglang.io/docs/DeepSeek/DeepSeek-V3_2)
- [Wan 2.1 GitHub](https://github.com/Wan-Video/Wan2.1)
- [SGLang Documentation](https://docs.sglang.ai/)
## ❓ Questions?
**Q: I only have consumer GPUs (4090/3090), can I help?**
A: Yes! You can help with documentation, examples, or testing the 1.3B Wan model. You can reach out @Richardczl98 for requesting additional GPUs
**Q: Which video model should we prioritize first?**
A: Wan 2.1 - it's the most mature open-source option.
**Q: Do I need to know SGLang internals?**
A: No, just familiarity with diffusion models and Python.
---
**Ready to contribute?** Drop a comment below! 🚀
cc @mickqian @Qiaolin-Yu @yhyang201 | https://github.com/sgl-project/sglang/issues/15443 | open | [] | 2025-12-19T03:44:33Z | 2025-12-23T13:09:31Z | 1 | Richardczl98 |
vllm-project/vllm | 30,969 | [Bug]: SmolLM3-3B FP8 Fails to Load [`compressed-tensors` and `transformers-impl` compatibility issue] | ### Your current environment
<details>
<summary>The output of <code>python collect_env.py</code></summary>
Running in official Docker image: vllm/vllm-openai:v0.11.1
GPU: NVIDIA L4 (GCP g2-standard-8)
`| NVIDIA-SMI 570.195.03 Driver Version: 570.195.03 CUDA Version: 12.9 |`
vLLM version: 0.11.1
```text
0.11.1
```
</details>
### 🐛 Describe the bug
vLLM v0.11.1 fails to load SmolLM3-3B FP8 quantized models with llm-compressor using compressed-tensors.
Same models work on v0.11.0.
Tested with:
- [huggingface.co/RedHatAI/SmolLM3-3B-FP8-dynamic](https://huggingface.co/RedHatAI/SmolLM3-3B-FP8-dynamic)
- Manually quantized fine tuned [SmolLM3-3B](https://huggingface.co/HuggingFaceTB/SmolLM3-3B) using llmcompressor==0.7 (compressed-tensors==0.12.2) in FP8-dynamic
- Manually quantized fine tuned [SmolLM3-3B(https://huggingface.co/HuggingFaceTB/SmolLM3-3B) using llmcompressor==0.8.1 (compressed-tensors==0.12.2) in FP8-dynamic
All fail on v0.11.1.
All work on v0.11.0.
Error occurs during model loading in find_matched_target function.
The error is: "Unable to find matching target for model.layers.0.self_attn.q_proj in the compressed-tensors config"
Complete error
```
+ exec python3 -m vllm.entrypoints.openai.api_server --model RedHatAI/SmolLM3-3B-FP8-dynamic --port 8000 --trust-remote-code --max-model-len 5000
[APIServer pid=1] INFO 12-12 05:05:29 [api_server.py:1772] vLLM API server version 0.11.1
[APIServer pid=1] INFO 12-12 05:05:29 [utils.py:253] non-default args: {'model': 'RedHatAI/SmolLM3-3B-FP8-dynamic', 'trust_remote_code': True, 'max_model_len': 5000}
[APIServer pid=1] The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored.
[APIServer pid=1] INFO 12-12 05:05:40 [model.py:637] Resolved architecture: SmolLM3ForCausalLM
[APIServer pid=1] INFO 12-12 05:05:40 [model.py:1750] Using max model len 5000
[APIServer pid=1] INFO 12-12 05:05:42 [scheduler.py:228] Chunked prefill is enabled with max_num_batched_tokens=2048.
[EngineCore_DP0 pid=37] INFO 12-12 05:05:54 [core.py:93] Initializing a V1 LLM engine (v0.11.1) with config: model='RedHatAI/SmolLM3-3B-FP8-dynamic', quantization=compressed-tensors
[EngineCore_DP0 pid=37] INFO 12-12 05:05:55 [parallel_state.py:1200] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://10.111.66.205:48123 backend=nccl
[EngineCore_DP0 pid=37] INFO 12-12 05:05:55 [parallel_state.py:1408] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank 0
[EngineCore_DP0 pid=37] INFO 12-12 05:05:55 [gpu_model_runner.py:3467] Starting to load model RedHatAI/SmolLM3-3B-FP8-dynamic...
[EngineCore_DP0 pid=37] INFO 12-12 05:05:56 [base.py:121] Using Transformers modeling backend.
[EngineCore_DP0 pid=37] ERROR 12-12 05:05:56 [core.py:843] EngineCore failed to start.
[EngineCore_DP0 pid=37] ERROR 12-12 05:05:56 [core.py:843] Traceback (most recent call last):
[EngineCore_DP0 pid=37] ERROR 12-12 05:05:56 [core.py:843] File "/usr/local/lib/python3.12/dist-packages/vllm/v1/engine/core.py", line 834, in run_engine_core
[EngineCore_DP0 pid=37] ERROR 12-12 05:05:56 [core.py:843] engine_core = EngineCoreProc(*args, **kwargs)
[EngineCore_DP0 pid=37] ERROR 12-12 05:05:56 [core.py:843] File "/usr/local/lib/python3.12/dist-packages/vllm/v1/engine/core.py", line 610, in __init__
[EngineCore_DP0 pid=37] ERROR 12-12 05:05:56 [core.py:843] File "/usr/local/lib/python3.12/dist-packages/vllm/v1/engine/core.py", line 102, in __init__
[EngineCore_DP0 pid=37] ERROR 12-12 05:05:56 [core.py:843] super().__init__(
[EngineCore_DP0 pid=37] ERROR 12-12 05:05:56 [core.py:843] File "/usr/local/lib/python3.12/dist-packages/vllm/v1/executor/abstract.py", line 101, in __init__
[EngineCore_DP0 pid=37] ERROR 12-12 05:05:56 [core.py:843] self.model_executor = executor_class(vllm_config)
[EngineCore_DP0 pid=37] ERROR 12-12 05:05:56 [core.py:843] File "/usr/local/lib/python3.12/dist-packages/vllm/v1/executor/uniproc_executor.py", line 48, in _init_executor
[EngineCore_DP0 pid=37] ERROR 12-12 05:05:56 [core.py:843] self._init_executor()
[EngineCore_DP0 pid=37] ERROR 12-12 05:05:56 [core.py:843] File "/usr/local/lib/python3.12/dist-packages/vllm/v1/worker/gpu_worker.py", line 273, in load_model
[EngineCore_DP0 pid=37] ERROR 12-12 05:05:56 [core.py:843] self.driver_worker.load_model()
[EngineCore_DP0 pid=37] ERROR 12-12 05:05:56 [core.py:843] File "/usr/local/lib/python3.12/dist-packages/vllm/v1/worker/gpu_model_runner.py", line 3484, in load_model
[EngineCore_DP0 pid=37] ERROR 12-12 05:05:56 [core.py:843] self.model_runner.load_model(eep_scale_up=eep_scale_up)
[EngineCore_DP0 pid=37] ERROR 12-12 05:05:56 [core.py:843] File "/usr/local/lib/python3.12/dist-packages/vllm/model_executor/model_loader/base_loader.py", line 49, in load_model
[EngineCore_DP0 pid=37] ERROR 12-12 05:05:56 [core.py:843] self.model = model_loader.load_model(
[EngineCore_DP0 pid=37] ERROR 12-12 05:05:56 [cor | https://github.com/vllm-project/vllm/issues/30969 | closed | [
"bug",
"help wanted",
"good first issue"
] | 2025-12-18T14:36:30Z | 2025-12-20T21:54:47Z | 3 | GauthierRoy |
huggingface/lerobot | 2,680 | Invalid frame index when training on merged datasets [RuntimeError] | ### Ticket Type
🐛 Bug Report (Something isn't working)
### Environment & System Info
```Shell
- LeRobot version: 0.4.3
- Platform: Linux-5.4.0-165-generic-x86_64-with-glibc2.35
- Python version: 3.10.12
- Huggingface Hub version: 0.35.3
- Datasets version: 4.1.1
- Numpy version: 2.2.6
- FFmpeg version: 4.4.2-0ubuntu0.22.04.1
- PyTorch version: 2.7.1+cu126
- Is PyTorch built with CUDA support?: True
- Cuda version: 12.6
- GPU model: Quadro RTX 6000
- Using GPU in script?: <fill in>
- lerobot scripts: ['lerobot-calibrate', 'lerobot-dataset-viz', 'lerobot-edit-dataset', 'lerobot-eval', 'lerobot-find-cameras', 'lerobot-find-joint-limits', 'lerobot-find-port', 'lerobot-imgtransform-viz', 'lerobot-info', 'lerobot-record', 'lerobot-replay', 'lerobot-setup-motors', 'lerobot-teleoperate', 'lerobot-train']
```
### Description
I'm having a problem when training a VLA with `lerobot-train` on a merged dataset.
I'm aware of the issue #2627 as well as PR #2550 that is supposed to fix the bug.
However, the problem is still occurring on the latest commit (4a151a9) of lerobot 0.4.3.
The dataset has been merged with the following script:
`lerobot-edit-dataset \
--repo_id whosricky/so101-megamix-v1 \
--operation.type merge \
--operation.repo_ids "['whosricky/so101_pick_red_cube_3cams', 'whosricky/so101_pick_blue_cube_3cams', 'whosricky/so101_pick_yellow_cube_3cams', 'whosricky/so101_pick_cube_reasoning_3cams', 'whosricky/so101_stacking_3cams', 'whosricky/so101_pickplace_red_cube_3cams', 'whosricky/so101_pickplace_all_red_cubes_3cams', 'whosricky/so101_sorting_cubes_3cams', 'whosricky/so101_pickplace_red_cubes_random_bowl_3cams']" \
--push_to_hub true `
Training on the single datasets works flawlessly. Training on the merged dataset results in an error.
The problematic sample seems to be #51 of "whosricky/so101_pick_blue_cube_3cams" due to the timestamp exceeding the default tolerance_s.
However, the problem occurs only on the merged dataset and not on the single one.
### Context & Reproduction
```
lerobot-train \
--dataset.repo_id=whosricky/so101-megamix-v1 \
--output_dir=outputs_xvla_megamix_v1/train/my_xvla \
--job_name=xvla_training_megamix_v1 \
--policy.path=lerobot/xvla-base \
--policy.repo_id=whosricky/xvla-so101-megamix-v1 \
--policy.private=true \
--policy.dtype=bfloat16 \
--num_workers=8 \
--batch_size=8 \
--steps=30000 \
--eval_freq=5000 \
--log_freq=100 \
--save_freq=5000 \
--policy.device=cuda \
--policy.freeze_vision_encoder=false \
--policy.freeze_language_encoder=false \
--policy.train_policy_transformer=true \
--policy.train_soft_prompts=true \
--policy.action_mode=auto \
--policy.num_image_views=3 \
--policy.empty_cameras=0 \
--rename_map='{"observation.images.top": "observation.images.image", "observation.images.gripper": "observation.images.image2", "observation.images.front": "observation.images.empty_camera_0"}' \
--wandb.enable=true
```
### Relevant logs or stack trace
```Shell
WARNING:accelerate.utils.other:Detected kernel version 5.4.0, which is below the recommended minimum of 5.5.0; this can cause the process to hang. It is recommended to upgrade the kernel to the minimum version or higher.
INFO 2025-12-18 12:38:22 ot_train.py:164 {'batch_size': 8,
'checkpoint_path': None,
'dataset': {'episodes': None,
'image_transforms': {'enable': False,
'max_num_transforms': 3,
'random_order': False,
'tfs': {'affine': {'kwargs': {'degrees': [-5.0,
5.0],
'translate': [0.05,
0.05]},
'type': 'RandomAffine',
'weight': 1.0},
'brightness': {'kwargs': {'brightness': [0.8,
1.2]},
'type': 'ColorJitter',
'weight': 1.0},
'contrast': {'kwargs': {'contrast': [0.8,
1.2]},
'type': 'ColorJitter',
'weight': 1.0},
'hue': {'kwargs': {'hue': [-0.05,
0.05]},
'type': 'ColorJitter',
'weight': 1.0},
'saturation': {'kwargs': {'satur | https://github.com/huggingface/lerobot/issues/2680 | open | [
"bug",
"question",
"dataset",
"visualization",
"examples",
"training"
] | 2025-12-18T13:29:50Z | 2025-12-26T06:26:37Z | null | RiccardoIzzo |
huggingface/trl | 4,719 | Loss calculation of `GKDTrainer` may be inaccurate when performing gradient accumulation? | It seems that `GKDTrainer` averages the loss of tokens in a micro batch ahead?
https://github.com/huggingface/trl/blob/8918c9836a3e0b43a6851c08d01b69072f56ca52/trl/experimental/gkd/gkd_trainer.py#L284 | https://github.com/huggingface/trl/issues/4719 | open | [
"🐛 bug",
"🏋 GKD"
] | 2025-12-18T12:50:05Z | 2025-12-18T12:50:49Z | 0 | jue-jue-zi |
huggingface/lerobot | 2,679 | Merging datasets removes fps from scalar features | ### Ticket Type
🐛 Bug Report (Something isn't working)
### Environment & System Info
```Shell
- LeRobot version: 0.4.3
- Platform: Linux-6.17.9-arch1-1-x86_64-with-glibc2.42
- Python version: 3.12.11
- Huggingface Hub version: 0.34.4
- Datasets version: 4.1.1
- Numpy version: 2.3.5
- FFmpeg version: n8.0.1
- PyTorch version: 2.7.1+cu128
- Is PyTorch built with CUDA support?: True
- Cuda version: 12.8
- GPU model: NVIDIA GeForce RTX 5090 Laptop GPU
- Using GPU in script?: <fill in>
- lerobot scripts: ['lerobot-calibrate', 'lerobot-dataset-viz', 'lerobot-edit-dataset', 'lerobot-eval', 'lerobot-find-cameras', 'lerobot-find-joint-limits', 'lerobot-find-port', 'lerobot-imgtransform-viz', 'lerobot-info', 'lerobot-record', 'lerobot-replay', 'lerobot-setup-motors', 'lerobot-teleoperate', 'lerobot-train']
```
### Description
When using the `merge_datasets` function, the fps attribute is removed from the scalar features in the dataset. Below are the scalar features from dataset.meta.features of a dataset before and after merging
Before:
```
'timestamp': {'dtype': 'float32', 'shape': (1,), 'names': None, 'fps': 10},
'frame_index': {'dtype': 'int64', 'shape': (1,), 'names': None, 'fps': 10},
'episode_index': {'dtype': 'int64', 'shape': (1,), 'names': None, 'fps': 10},
'index': {'dtype': 'int64', 'shape': (1,), 'names': None, 'fps': 10},
'task_index': {'dtype': 'int64', 'shape': (1,), 'names': None, 'fps': 10}}
```
After:
```
'timestamp': {'dtype': 'float32', 'shape': (1,), 'names': None},
'frame_index': {'dtype': 'int64', 'shape': (1,), 'names': None},
'episode_index': {'dtype': 'int64', 'shape': (1,), 'names': None},
'index': {'dtype': 'int64', 'shape': (1,), 'names': None},
'task_index': {'dtype': 'int64', 'shape': (1,), 'names': None}
```
This creates subsequent problems when trying to add an additional dataset to a merged output as the feature mismatch will cause an error to be thrown
### Context & Reproduction
Running the script below shows the features change before and after the merge
```
from lerobot.datasets.dataset_tools import split_dataset, merge_datasets
from lerobot.datasets.lerobot_dataset import LeRobotDataset
from pprint import pprint
dataset = LeRobotDataset("lerobot/pusht")
feat_1 = dataset.meta.features
splits = split_dataset(dataset, splits={"train": 0.8, "val": 0.2})
merged = merge_datasets([splits["train"], splits["val"]], output_repo_id="lerobot/pusht_merged")
feat_2 = merged.meta.features
print("Features of original dataset:")
pprint(feat_1)
print("Features of merged dataset:")
pprint(feat_2)
```
### Relevant logs or stack trace
```Shell
Features of original dataset:
{'action': {'dtype': 'float32',
'fps': 10.0,
'names': {'motors': ['motor_0', 'motor_1']},
'shape': (2,)},
'episode_index': {'dtype': 'int64', 'fps': 10.0, 'names': None, 'shape': (1,)},
'frame_index': {'dtype': 'int64', 'fps': 10.0, 'names': None, 'shape': (1,)},
'index': {'dtype': 'int64', 'fps': 10.0, 'names': None, 'shape': (1,)},
'next.done': {'dtype': 'bool', 'fps': 10.0, 'names': None, 'shape': (1,)},
'next.reward': {'dtype': 'float32', 'fps': 10.0, 'names': None, 'shape': (1,)},
'next.success': {'dtype': 'bool', 'fps': 10.0, 'names': None, 'shape': (1,)},
'observation.image': {'dtype': 'video',
'names': ['height', 'width', 'channel'],
'shape': (96, 96, 3),
'video_info': {'has_audio': False,
'video.codec': 'av1',
'video.fps': 10.0,
'video.is_depth_map': False,
'video.pix_fmt': 'yuv420p'}},
'observation.state': {'dtype': 'float32',
'fps': 10.0,
'names': {'motors': ['motor_0', 'motor_1']},
'shape': (2,)},
'task_index': {'dtype': 'int64', 'fps': 10.0, 'names': None, 'shape': (1,)},
'timestamp': {'dtype': 'float32', 'fps': 10.0, 'names': None, 'shape': (1,)}}
Features of merged dataset:
{'action': {'dtype': 'float32',
'fps': 10.0,
'names': {'motors': ['motor_0', 'motor_1']},
'shape': (2,)},
'episode_index': {'dtype': 'int64', 'names': None, 'shape': (1,)},
'frame_index': {'dtype': 'int64', 'names': None, 'shape': (1,)},
'index': {'dtype': 'int64', 'names': None, 'shape': (1,)},
'next.done': {'dtype': 'bool', 'fps': 10.0, 'names': None, 'shape': (1,)},
'next.reward': {'dtype': 'float32', 'fps': 10.0, 'names': None, 'shape': (1,)},
'next.success': {'dtype': 'bool', 'fps': 10.0, 'names': None, 'shape': (1,)},
'observation.image': {'dtype': 'video',
'names': ['height', 'width', 'channel'],
'shape': (96, 96, 3),
'video_info': {'has_audio': False,
'video.codec': 'av1',
'video.fps': 10.0,
| https://github.com/huggingface/lerobot/issues/2679 | open | [
"bug",
"enhancement",
"question",
"dataset",
"performance",
"examples"
] | 2025-12-18T12:47:14Z | 2025-12-18T15:25:12Z | null | reeceomahoney |
vllm-project/vllm | 30,956 | [Feature]: could output the given format logger ? | ### 🚀 The feature, motivation and pitch
hi,dear ,
i have def the logger from py scripts ,etc, logger_utils.py
and could i use shell run the command with the logger,
such as ,
`vllm serve qwen3-embedding-0.6b --logger_file logger_utils.py `
thx
i really need your help
SOS ,thx
### Alternatives
_No response_
### Additional context
_No response_
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/30956 | open | [
"feature request"
] | 2025-12-18T09:35:22Z | 2025-12-19T01:52:41Z | 5 | ucas010 |
huggingface/lerobot | 2,678 | Bug: lerobot-dataset-viz IndexError when visualizing specific episodes | # Bug Report: `lerobot-dataset-viz` IndexError when visualizing specific episodes
## Description
The `lerobot-dataset-viz` command fails with an `IndexError` when trying to visualize a specific episode using the `--episode-index` parameter. The issue is caused by `EpisodeSampler` using global dataset indices while the dataset has been filtered to contain only the specified episode.
## Error Message
```
IndexError: Invalid key: 180 is out of bounds for size 180
```
Full traceback:
```
Traceback (most recent call last):
File "/path/to/lerobot/scripts/lerobot_dataset_viz.py", line 289, in main
visualize_dataset(dataset, **vars(args))
File "/path/to/lerobot/scripts/lerobot_dataset_viz.py", line 148, in visualize_dataset
for batch in tqdm.tqdm(dataloader, total=len(dataloader)):
...
File "/path/to/lerobot/datasets/lerobot_dataset.py", line 1028, in __getitem__
item = self.hf_dataset[idx]
...
IndexError: Invalid key: 180 is out of bounds for size 180
```
## Steps to Reproduce
1. Create a LeRobot dataset with multiple episodes (e.g., 20 episodes, 180 frames each)
2. Try to visualize episode 1:
```bash
lerobot-dataset-viz \
--repo-id lerobot/test \
--root ./lerobot_dataset \
--mode local \
--episode-index 1 \
--batch-size 2
```
3. Error occurs when trying to load the data
## Root Cause Analysis
The bug is in the `EpisodeSampler` class (line 81-91 of `lerobot_dataset_viz.py`):
```python
class EpisodeSampler(torch.utils.data.Sampler):
def __init__(self, dataset: LeRobotDataset, episode_index: int):
from_idx = dataset.meta.episodes["dataset_from_index"][episode_index] # 180
to_idx = dataset.meta.episodes["dataset_to_index"][episode_index] # 360
self.frame_ids = range(from_idx, to_idx) # range(180, 360)
```
**The problem:**
1. At line 287, the dataset is filtered: `dataset = LeRobotDataset(repo_id, episodes=[args.episode_index], ...)`
2. The filtered dataset only contains 180 frames with **local indices 0-179**
3. But `EpisodeSampler` uses indices from `dataset.meta.episodes` which are **global indices 180-359** (position in the full dataset)
4. When DataLoader tries to access `dataset[180]`, it fails because the filtered dataset only has indices 0-179
**Example:**
```
Full dataset (3600 frames):
┌──────────┬──────────┬──────────┬─────┬──────────┐
│ Episode 0│ Episode 1│ Episode 2│ ... │ Episode 19│
│ 0-179 │ 180-359 │ 360-539 │ ... │ 3420-3599│
└──────────┴──────────┴──────────┴─────┴──────────┘
↑
Global indices
Filtered dataset (180 frames, episode 1 only):
┌──────────┐
│ Episode 1│ ← Only these 180 frames exist
│ 0-179 │ ← Local indices in filtered dataset
└──────────┘
EpisodeSampler tries to use: range(180, 360) ✗ Out of bounds!
```
## Proposed Fix
Modify `EpisodeSampler` to handle filtered datasets:
```python
class EpisodeSampler(torch.utils.data.Sampler):
def __init__(self, dataset: LeRobotDataset, episode_index: int):
# Check if dataset is already filtered to a single episode
if dataset.episodes is not None and len(dataset.episodes) == 1:
# Dataset is filtered, use all available frames (local indices)
self.frame_ids = range(len(dataset))
else:
# Dataset is not filtered, use global indices from metadata
from_idx = dataset.meta.episodes["dataset_from_index"][episode_index]
to_idx = dataset.meta.episodes["dataset_to_index"][episode_index]
self.frame_ids = range(from_idx, to_idx)
def __iter__(self) -> Iterator:
return iter(self.frame_ids)
def __len__(self) -> int:
return len(self.frame_ids)
```
## Workaround
Until this is fixed, users can visualize a specific episode by:
1. Loading the full dataset without filtering
2. Using `torch.utils.data.Subset` to select the episode
```python
import rerun as rr
from lerobot.datasets.lerobot_dataset import LeRobotDataset
from torch.utils.data import DataLoader, Subset
# Load full dataset (no filtering)
dataset = LeRobotDataset(
repo_id="lerobot/test",
root="./lerobot_dataset"
)
# Manually select episode frames
episode_index = 1
from_idx = dataset.meta.episodes[episode_index]["dataset_from_index"]
to_idx = dataset.meta.episodes[episode_index]["dataset_to_index"]
episode_dataset = Subset(dataset, range(from_idx, to_idx))
# Create dataloader
dataloader = DataLoader(episode_dataset, batch_size=2, shuffle=False)
# Visualize...
```
## Environment
- **LeRobot Version:** 0.4.2
- **Python Version:** 3.12.11
- **PyTorch Version:** 2.7.1+cu126
- **Datasets Version:** 4.1.1
- **OS:** Linux
## Additional Context
This issue affects any dataset where users want to visualize a specific episode that is not episode 0. The bug makes the `--episode-index` parameter effectively unusable for episodes other than the first one when the dataset has already been filtered.
## Impact
- **Severity:** Medium (cor | https://github.com/huggingface/lerobot/issues/2678 | open | [
"bug",
"question",
"dataset",
"visualization",
"python",
"examples"
] | 2025-12-18T08:45:05Z | 2025-12-24T08:31:00Z | null | apeSh1t |
vllm-project/vllm | 30,941 | [Performance]: Why Does Latency Remain Unchanged in vLLM 0.11.0 When Input Token Count Decreases for qwen3-vl-30b-a3b? | ### Proposal to improve performance
_No response_
### Report of performance regression
_No response_
### Misc discussion on performance
Using vLLM version 0.11.0 to run the qwen3-vl-30b-a3b model, the stress test results show that although the number of input tokens decreases, the latency does not change.
The model is deployed on a single A800 GPU. The startup command is:
vllm server
--dtype bfloat16
--max-model-len 128000
--gpu-memory-utilization 0.95
--limit-mm-per-prompt.video 0
I performed a stress test using one image and a set of text prompts, with QPS set to 10.
I resized the image to 0.25x and 0.7x of the original size while keeping everything else unchanged.
The conclusions are as follows:
qwen3-30b-a3b (single image *0.25) latency 3s
qwen3-30b-a3b (single image *0.7) latency 5s
qwen3-30b-a3b (single image) latency 5s
Prior conditions:
Input token scale / Output token scale
Single image + text prompts: about 4200 / about 70
Single image *0.6 + text prompts: about 1900 / about 70
Single image *0.3 + text prompts: about 860 / about 70
### Your current environment (if you think it is necessary)
```text
The output of `python collect_env.py`
```
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/30941 | open | [
"performance"
] | 2025-12-18T07:40:35Z | 2025-12-18T07:40:35Z | 0 | Hormoney |
vllm-project/vllm | 30,933 | [Usage]: What is the latest instruction to run DeepSeek V3.2? | ### Your current environment
vLLM 0.12.0
### How would you like to use vllm
I am following the guidelines here https://docs.vllm.ai/projects/recipes/en/latest/DeepSeek/DeepSeek-V3_2.html for running DeepSeek v3.2. By following the instructions I installed vLLM 0.12.0 on my H200 node. However, when I try to run it with `vllm serve deepseek-ai/DeepSeek-V3.2 --tensor-parallel-size 8 --tokenizer-mode deepseek_v32` it gives an error
```
(APIServer pid=816209) ValueError: No tokenizer registered for tokenizer_mode='deepseek_v32'.
```
If I do not include the `--tokenizer-mode` then the server spins up with no errors, but when I try to send a request, I get another error below
```
(APIServer pid=753941) ERROR 12-18 06:04:47 [serving_chat.py:263] ValueError: As of transformers v4.44, default chat template is no longer allowed, so you must provide a chat template if the tokenizer does not define one.
```
I am wondering if there is an update on the instructions to run DeepSeek V3.2 on vLLM.
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/30933 | open | [
"usage"
] | 2025-12-18T06:18:29Z | 2025-12-18T15:50:29Z | 1 | IKACE |
vllm-project/vllm | 30,923 | [Bug]: Use the offical doucment vllm online method deploy DeepSeek-OCR，the result is very bad . but I ust the offline method the result is normal. why ? | ### Your current environment
<details>
<summary>The output of <code>python collect_env.py</code></summary>
```text
Your output of `python collect_env.py` here
```
</details>
### 🐛 Describe the bug
I use https://github.com/vllm-project/recipes/blob/main/DeepSeek/DeepSeek-OCR.md
the offline and online mehtod is work, run ok。
but the same picture in offline is better than online, I can't find the reason what happend ? can someone help me
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/30923 | closed | [
"bug"
] | 2025-12-18T04:14:33Z | 2025-12-18T04:25:20Z | 0 | git-liweichao |
vllm-project/vllm | 30,922 | [Bug]: Use the offical doucment vllm online method deploy DeepSeek-OCR，the result is very bad . but I ust the offline method the result is normal. why ? | ### Your current environment
<details>
<summary>The output of <code>python collect_env.py</code></summary>
```text
Your output of `python collect_env.py` here
```
</details>
### 🐛 Describe the bug
I use https://github.com/vllm-project/recipes/blob/main/DeepSeek/DeepSeek-OCR.md
the offline and online mehtod is work, run ok。
but the same picture in offline is better than online, I can't find the reason what happend ? can someone help me
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/30922 | open | [
"bug"
] | 2025-12-18T04:08:46Z | 2025-12-18T04:25:36Z | 1 | git-liweichao |
sgl-project/sglang | 15,359 | [Bug] The handling logic for tool_choice = 'auto' in the DeepseekV3.2 model may be incorrect. | ### Checklist
- [ ] I searched related issues but found no solution.
- [ ] The bug persists in the latest version.
- [ ] Issues without environment info and a minimal reproducible demo are hard to resolve and may receive no feedback.
- [ ] If this is not a bug report but a general question, please start a discussion at https://github.com/sgl-project/sglang/discussions. Otherwise, it will be closed.
- [ ] Please use English. Otherwise, it will be closed.
### Describe the bug
When using SGLang (sglang:v0.5.6.post2) with DeepseekV3.2, I noticed the response of some request which involves tool calls is not currect.
like the following requests
```sh
curl -X POST http://{host}:{port}/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer sk-1234" \
-d '{
"model": "DeepseekV3.2",
"messages": [
{
"role": "user",
"content": "What is the weather in Beijing?"
}
],
"tools": [
{
"type": "function",
"function": {
"name": "get_current_weather",
"description": "Get the current weather in a given location",
"strict": true,
"parameters": {
"type": "object",
"properties": {
"location": {
"type": "string",
"description": "The city and state, e.g. San Francisco, CA"
},
"unit": {
"type": "string",
"enum": ["celsius", "fahrenheit"]
}
},
"required": ["location"]
}
}
}
],
"tool_choice": "auto",
"stream": false
}'
```
might response something like
```sh
{"id":"88c2a168ad43446f9116aeed715cd835","object":"chat.completion","created":1766024807,"model":"DeepseekV3.2","choices":[{"index":0,"message":{"role":"assistant","content":"tool_call_name=current_weather","reasoning_content":null,"tool_calls":null},"logprobs":null,"finish_reason":"stop","matched_stop":1}],"usage":{"prompt_tokens":198,"total_tokens":206,"completion_tokens":8,"prompt_tokens_details":null,"reasoning_tokens":0},"metadata":{"weight_version":"default"}}
```
or
```sh
{"id":"0223b02af05b4c9b99e8b9e4b2abab12","object":"chat.completion","created":1766026261,"model":"DeepseekV3.2","choices":[{"index":0,"message":{"role":"assistant","content":"tool_call_name: get_current_weather\ntool_call_arguments: {\n \"location\": \"Beijing, China\",\n \"unit\": \"celsius\"\n}","reasoning_content":null,"tool_calls":null},"logprobs":null,"finish_reason":"stop","matched_stop":1}],"usage":{"prompt_tokens":198,"total_tokens":232,"completion_tokens":34,"prompt_tokens_details":null,"reasoning_tokens":0},"metadata":{"weight_version":"default"}}
```
As you can see from the response, the content value contains `tool_call_name` but tool_calls is set to `null`
And if change tool_choice to 'required', the response looks like
```sh
{"id":"550109a7f6854af3ba47fdad4f38f9d5","object":"chat.completion","created":1766025639,"model":"DeepseekV3.2","choices":[{"index":0,"message":{"role":"assistant","content":null,"reasoning_content":null,"tool_calls":[{"id":"call_f171fbf82d7d41dab0eaf258","index":0,"type":"function","function":{"name":"get_current_weather","arguments":"{\"location\": \"Beijing, China\", \"unit\": \"celsius\"}"}}]},"logprobs":null,"finish_reason":"tool_calls","matched_stop":null}],"usage":{"prompt_tokens":198,"total_tokens":229,"completion_tokens":31,"prompt_tokens_details":null,"reasoning_tokens":0},"metadata":{"weight_version":"default"}}
```
and when checking the source codes, I find it might be related to the following codes
https://github.com/sgl-project/sglang/blob/9e7656be80578fe981a723bd115373371a9d0d90/python/sglang/srt/entrypoints/openai/serving_chat.py#L248-L260
https://github.com/sgl-project/sglang/blob/9e7656be80578fe981a723bd115373371a9d0d90/python/sglang/srt/function_call/function_call_parser.py#L189-L201
### Reproduction
start SGLang with the following command
```sh
python3 -m sglang.launch_server --model /root/.cache/huggingface/DeepSeek-V3.2 --served-model-name VILLM-N2 --tp 8 --ep 8 --dp 8 --enable-dp-attention --trust-remote-code --port 30000 --host 0.0.0.0 --enable-metrics --mem-fraction-static 0.75 --cuda-graph-max-bs 128 --torch-compile-max-bs 8 --speculative-algorithm EAGLE --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 --nsa-prefill-backend flashmla_sparse --nsa-decode-backend fa3 --grammar-backend xgrammar --reasoning-parser deepseek-v3 --tool-call-parser deepseekv32 --chat-template ./examples/chat_template/tool_chat_template_deepseekv32.jinja
```
send request
```sh
curl -X POST http://{host}:{port}/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer sk-1234" \
-d '{
"model": "DeepseekV3.2",
"messages": [
{
"role": "user",
"content": "What is the weather in Beijing?"
}
],
"tools": [
| https://github.com/sgl-project/sglang/issues/15359 | closed | [] | 2025-12-18T02:47:26Z | 2025-12-18T03:36:38Z | 4 | JerryKwan |
huggingface/lerobot | 2,673 | Dataset v2 not working anymore | ### Ticket Type
Feature
### Environment & System Info
```Shell
- LeRobot version: 0.4.3
- Platform: macOS-26.2-arm64-arm-64bit
- Python version: 3.10.19
- Huggingface Hub version: 0.35.3
- Datasets version: 4.1.1
- Numpy version: 2.2.6
- FFmpeg version: 7.1.1
- PyTorch version: 2.7.1
- Is PyTorch built with CUDA support?: False
- Cuda version: N/A
- GPU model: N/A
- Using GPU in script?: <fill in>
- lerobot scripts: ['lerobot-calibrate', 'lerobot-dataset-viz', 'lerobot-edit-dataset', 'lerobot-eval', 'lerobot-find-cameras', 'lerobot-find-joint-limits', 'lerobot-find-port', 'lerobot-imgtransform-viz', 'lerobot-info', 'lerobot-record', 'lerobot-replay', 'lerobot-setup-motors', 'lerobot-teleoperate', 'lerobot-train']
```
### Description
I did git pull and my dataset v2 doesn't work anymore. My model raises with the logs below.
### Context & Reproduction
1. `lerobot-train --help`
2. Check outputs
### Relevant logs or stack trace
```Shell
File "/admin/home/michel_aratingi/miniconda3/envs/groot/lib/python3.10/site-packages/torch/utils/data/dataloader.py", line 733, in __next__
data = self._next_data()
File "/admin/home/michel_aratingi/miniconda3/envs/groot/lib/python3.10/site-packages/torch/utils/data/dataloader.py", line 1488, in _next_data
return self._process_data(data, worker_id)
File "/admin/home/michel_aratingi/miniconda3/envs/groot/lib/python3.10/site-packages/torch/utils/data/dataloader.py", line 1550, in _process_data
data.reraise()
File "/admin/home/michel_aratingi/miniconda3/envs/groot/lib/python3.10/site-packages/torch/_utils.py", line 750, in reraise
raise exception
IndexError: Caught IndexError in DataLoader worker process 1.
Original Traceback (most recent call last):
File "/admin/home/michel_aratingi/miniconda3/envs/groot/lib/python3.10/site-packages/torch/utils/data/_utils/worker.py", line 349, in _worker_loop
data = fetcher.fetch(index) # type: ignore[possibly-undefined]
File "/admin/home/michel_aratingi/miniconda3/envs/groot/lib/python3.10/site-packages/torch/utils/data/_utils/fetch.py", line 52, in fetch
data = [self.dataset[idx] for idx in possibly_batched_index]
File "/admin/home/michel_aratingi/miniconda3/envs/groot/lib/python3.10/site-packages/torch/utils/data/_utils/fetch.py", line 52, in <listcomp>
data = [self.dataset[idx] for idx in possibly_batched_index]
File "/admin/home/michel_aratingi/code/collab-lerobot/src/lerobot/datasets/lerobot_dataset.py", line 975, in __getitem__
item = self.hf_dataset[idx]
File "/admin/home/michel_aratingi/miniconda3/envs/groot/lib/python3.10/site-packages/datasets/arrow_dataset.py", line 2859, in __getitem__
return self._getitem(key)
File "/admin/home/michel_aratingi/miniconda3/envs/groot/lib/python3.10/site-packages/datasets/arrow_dataset.py", line 2840, in _getitem
pa_subtable = query_table(self._data, key, indices=self._indices)
File "/admin/home/michel_aratingi/miniconda3/envs/groot/lib/python3.10/site-packages/datasets/formatting/formatting.py", line 612, in query_table
_check_valid_index_key(key, size)
File "/admin/home/michel_aratingi/miniconda3/envs/groot/lib/python3.10/site-packages/datasets/formatting/formatting.py", line 552, in _check_valid_index_key
raise IndexError(f"Invalid key: {key} is out of bounds for size {size}")
IndexError: Invalid key: 46969 is out of bounds for size 46963
```
### Checklist
- [x] I have searched existing tickets to ensure this isn't a duplicate.
- [x] I am using the latest version of the `main` branch.
- [x] (I have verified this is not an environment-specific problem.
### Additional Info / Workarounds
Maybe if I try to update my transformers dependency?
I edit this ticket | https://github.com/huggingface/lerobot/issues/2673 | closed | [
"enhancement",
"question",
"dataset",
"dependencies",
"training"
] | 2025-12-17T21:35:31Z | 2025-12-17T23:26:54Z | null | imstevenpmwork |
huggingface/lerobot | 2,670 | Async inference for simulation (libero benchmark) | ### Issue Type
{"label" => "❓ Technical Question"}
### Environment & System Info
```Shell
```
### Description
Is there any way that we can support async inference for simulator (e.g., libero)? This makes it possible to test RTC with simulators.
### Context & Reproduction
A question re a feature.
### Expected Behavior / Desired Outcome
_No response_
### Relevant logs or stack trace
```Shell
```
### Checklist
- [ ] I have searched existing issues to ensure this isn't a duplicate.
- [ ] I am using the latest version of the `main` branch.
- [ ] (For bugs) I have verified this is not an environment-specific issue.
### Additional Info / Workarounds
_No response_ | https://github.com/huggingface/lerobot/issues/2670 | open | [
"question",
"simulation",
"performance",
"evaluation"
] | 2025-12-17T18:57:07Z | 2026-01-02T05:40:18Z | null | dywsjtu |
huggingface/transformers | 42,930 | Inconsistent handling of video_metadata in Qwen3VLVideoProcessor usage example | ### System Info
transformers==4.57.3
### Who can help?
@zucchini-nlp @yonigozlan @molbap
### Information
- [x] The official example scripts
- [ ] My own modified scripts
### Tasks
- [ ] An officially supported task in the `examples` folder (such as GLUE/SQuAD, ...)
- [x] My own task or dataset (give details below)
### Reproduction
I'm working with the `Qwen3VLVideoProcessor` and noticed a potential inconsistency between the processor's output and its expected usage.
According to the current implementation of `Qwen3VLVideoProcessor._preprocess()`, the returned `BatchFeature` only contains the keys:
- `"pixel_values_videos"`
- `"video_grid_thw"`
However, in some calling code, I see logic like:
```python
videos_inputs = self.video_processor(videos=videos, **kwargs)
if "return_metadata" not in kwargs:
video_metadata = videos_inputs.pop("video_metadata")
```
How does it work? thank you very much
### Expected behavior
I want to change Qwen2.5vl to Qwen3vl but can't set a fixed nframes | https://github.com/huggingface/transformers/issues/42930 | closed | [
"bug"
] | 2025-12-17T17:21:00Z | 2025-12-18T10:32:23Z | 3 | wagoriginal |
vllm-project/vllm | 30,882 | [Bug]: Marlin Fp8 Block Quant Failure | ### Your current environment
<details>
<summary>The output of <code>python collect_env.py</code></summary>
```text
Your output of `python collect_env.py` here
```
</details>
### 🐛 Describe the bug
```bash
MODEL := "Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8"
#MODEL := "RedHatAI/Mixtral-8x7B-Instruct-v0.1-FP8"
launch_marlin:
VLLM_TEST_FORCE_FP8_MARLIN=1 VLLM_USE_DEEPGEMM=0 chg run --gpus 1 -- vllm serve {{MODEL}} --enforce-eager --max-model-len 8192
eval:
lm_eval \
--model local-completions \
--tasks gsm8k \
--model_args "model={{MODEL}},base_url=http://localhost:8000/v1/completions,num_concurrent=1000,tokenized_requests=False"
```
Result:
```bash
(vllm) [robertgshaw2-redhat@nm-automation-h100-standalone-1-preserve vllm]$ just launch_marlin
VLLM_TEST_FORCE_FP8_MARLIN=1 VLLM_USE_DEEPGEMM=0 chg run --gpus 1 -- vllm serve Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8 --enforce-eager --max-model-len 8192
Reserved 1 GPU(s): [1] for command execution
(APIServer pid=3634068) INFO 12-17 15:54:23 [api_server.py:1259] vLLM API server version 0.13.0rc2.dev185+g00a8d7628
(APIServer pid=3634068) INFO 12-17 15:54:23 [utils.py:253] non-default args: {'model_tag': 'Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8', 'model': 'Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8', 'max_model_len': 8192, 'enforce_eager': True}
(APIServer pid=3634068) INFO 12-17 15:54:23 [model.py:514] Resolved architecture: Qwen3MoeForCausalLM
(APIServer pid=3634068) INFO 12-17 15:54:23 [model.py:1661] Using max model len 8192
(APIServer pid=3634068) INFO 12-17 15:54:24 [scheduler.py:230] Chunked prefill is enabled with max_num_batched_tokens=8192.
(APIServer pid=3634068) WARNING 12-17 15:54:24 [vllm.py:622] Enforce eager set, overriding optimization level to -O0
(APIServer pid=3634068) INFO 12-17 15:54:24 [vllm.py:722] Cudagraph is disabled under eager mode
(EngineCore_DP0 pid=3634329) INFO 12-17 15:54:31 [core.py:93] Initializing a V1 LLM engine (v0.13.0rc2.dev185+g00a8d7628) with config: model='Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8', speculative_config=None, tokenizer='Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=8192, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, disable_custom_all_reduce=False, quantization=fp8, enforce_eager=True, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False), seed=0, served_model_name=Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'level': None, 'mode': <CompilationMode.NONE: 0>, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['+quant_fp8', 'all', '+quant_fp8'], 'splitting_ops': [], 'compile_mm_encoder': False, 'compile_sizes': [], 'compile_ranges_split_points': [8192], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': <CUDAGraphMode.NONE: 0>, 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': False, 'fuse_act_quant': False, 'fuse_attn_quant': False, 'eliminate_noops': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': <DynamicShapesType.BACKED: 'backed'>, 'evaluate_guards': False}, 'local_cache_dir': None}
(EngineCore_DP0 pid=3634329) INFO 12-17 15:54:32 [parallel_state.py:1210] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://10.243.64.5:43323 backend=nccl
(EngineCore_DP0 pid=3634329) INFO 12-17 15:54:32 [parallel_state.py:1418] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank 0
(EngineCore_DP0 pid=3634329) INFO 12-17 15:54:33 [gpu_model_runner.py:3620] Starting to load model Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8...
(EngineCore_DP0 pid=3634329) INFO 12-17 15:54:33 [deep_gemm.py:76] DeepGEMM E8M0 enabled on current platform.
(EngineCore_DP0 pid=3634329) INFO 12-17 15:54:33 [cuda.py:351] Using FLASH_ATTN attention backend out of potential backends: ('FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION')
(EngineCore_DP0 pid=3634329) INFO 12-17 15:54:33 [layer.py:373] Enabled separate cuda str | https://github.com/vllm-project/vllm/issues/30882 | closed | [
"bug",
"help wanted",
"good first issue"
] | 2025-12-17T15:55:18Z | 2025-12-17T16:02:54Z | 2 | robertgshaw2-redhat |
vllm-project/vllm | 30,879 | [Doc]: Add some documentation about encoder compilation | ### 📚 The doc issue
I want something like a design doc for encoder compilation. For example:
- It uses support_torch_compile and set_model_tag to avoid cache collisions
- it supports or doesn't support the following features that VllmBackend does: cudagraphs, compile_ranges, and a high-level explanation for how these are turned off or on.
- it inherits from compilation_config (or maybe it doesn't)
- here's how to turn it on/off
I'm having a difficult time thinking through the edge cases in https://github.com/vllm-project/vllm/pull/30822 and https://github.com/vllm-project/vllm/pull/30489
cc @Lucaskabela
### Suggest a potential alternative/fix
_No response_
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/30879 | open | [
"documentation",
"torch.compile"
] | 2025-12-17T15:44:50Z | 2025-12-17T16:27:38Z | 1 | zou3519 |
vllm-project/vllm | 30,865 | [Usage]:Tools GLM4.6v with vLLM | ### Your current environment
Hello,
I am running tests on this model, which I find excellent. However, I am encountering a few issues and would like to know whether it is possible to fix them or if I am simply asking for the impossible.
First of all, here is my vLLM configuration:
`docker run -d \ --name vllm-llm \ --gpus '"device=4,5,6,7"' \ -e NVIDIA_DRIVER_CAPABILITIES=compute,utility \ -e VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME="${SHM_NAME}" \ -v /raid/workspace/qladane/vllm/hf-cache:/root/.cache/huggingface \ --env "HF_TOKEN=${HF_TOKEN:-}" \ -p 8003:8000 \ --ipc=host \ --restart unless-stopped \ vllm-openai:glm46v \ zai-org/GLM-4.6V-FP8 \ --tensor-parallel-size 4 \ --enforce-eager \ --served-model-name ImagineAI \ --allowed-local-media-path / \ --limit-mm-per-prompt '{"image": 1, "video": 0}' \ --max-model-len 131072 \ --dtype auto \ --kv-cache-dtype fp8 \ --gpu-memory-utilization 0.85 \ --reasoning-parser glm45 \ --tool-call-parser glm45 \ --enable-auto-tool-choice \ --enable-expert-parallel \ --mm-encoder-tp-mode data \ --mm-processor-cache-type shm`
Next, here is my OpenWebUI configuration:
<img width="1080" height="568" alt="Image" src="https://github.com/user-attachments/assets/af5ff9c0-9cdc-407f-8b0b-8e76a42746af" />
<img width="1080" height="394" alt="Image" src="https://github.com/user-attachments/assets/60fa32f9-2f54-4a75-8dc1-0ed00c69c4e5" />
<img width="1080" height="416" alt="Image" src="https://github.com/user-attachments/assets/783ba2e7-08e9-426a-a8be-9a2a561b2fe0" />
<img width="1080" height="357" alt="Image" src="https://github.com/user-attachments/assets/a7c30850-a680-401a-b149-5787000e7344" />
I would like to know whether, with GLM-4.6V and OpenWebUI, it is possible to make the model choose and execute tools autonomously when it considers them relevant.
At the moment:
If it is an internet search, I have to manually activate the button, even though access is already available.
If it is Python code, I have to click “execute”; it does not run it by itself, even though it clearly has access to Jupyter, etc.
If anyone has already encountered this issue.
Thank you very much in advance for your help.
Kind regards
### How would you like to use vllm
I want to run inference of a [specific model](put link here). I don't know how to integrate it with vllm.
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/30865 | open | [
"usage"
] | 2025-12-17T10:51:34Z | 2025-12-18T08:33:44Z | 1 | qBrabus |
sgl-project/sglang | 15,321 | [Feature][VLM] Support ViT Piecewise CUDA Graph for VLMs | ### Checklist
- [ ] If this is not a feature request but a general question, please start a discussion at https://github.com/sgl-project/sglang/discussions. Otherwise, it will be closed.
- [ ] Please use English. Otherwise, it will be closed.
### Motivation
Support ViT Piecewise CUDA Graph for VLMs can improve prefill performance for VLMs.
- [x] Support ViT PCG Framework https://github.com/sgl-project/sglang/pull/14422
- [x] Support Qwen2.5-VL https://github.com/sgl-project/sglang/pull/14422
- [x] Support Qwen3-VL https://github.com/sgl-project/sglang/pull/15320
- [ ] Support InternVL
- [ ] Support GLM-4.1V
### Related resources
_No response_ | https://github.com/sgl-project/sglang/issues/15321 | open | [
"performance",
"Multi-modal",
"vlm"
] | 2025-12-17T09:17:18Z | 2026-01-04T02:09:13Z | 0 | yuan-luo |
vllm-project/vllm | 30,859 | [Bug]: set_current_vllm_config() is only done during the initialization stage but not the runtime stage | ### Your current environment
Any env
### 🐛 Describe the bug
# Issue Statement
Currently, `set_current_vllm_config()` is only done during the initialization stage but not the runtime stage. If the code tries to call `get_current_vllm_config()`, vLLM prints a warning "Current vLLM config is not set." and returns a default config.
However, this approach is problematic because:
1. When contributors change the code, many of us did not realize the fact that `get_current_vllm_config()` should only be called during init stage and should not be called during runtime stage.
2. It's just a warning instead of a hard failure, so contributors may not notice this when they run local tests.
3. Such warnings could be annoying to users because it may be printed for every single decoding step. Plus, the warning doesn't carry any useful info about how to fix/bypass the issue.
4. The default config may be completely incorrect for the caller function.
5. Warning prints on every step might impact performance, because print isn't fast operation. (thanks to @vadiklyutiy )
# Requirements
We should change the behavior such that:
- `get_current_vllm_config()` either returns the real config set by the user or raises an error if the config does not exist.
# Related Issues
This issues have appeared many times in the past. Although the fix is usually not difficult, it is an annoying recurrent issues that we should avoid in the future to avoid wasted engineering effort.
- https://github.com/vllm-project/vllm/issues/13207
- https://github.com/vllm-project/vllm/pull/29999
- https://github.com/vllm-project/vllm/issues/30185
- https://github.com/vllm-project/vllm/issues/30240
- https://github.com/vllm-project/vllm/issues/30571
# Possible Solutions
## Solution A: `set_current_vllm_config()` for runtime stage as well
Such that `get_current_vllm_config()` is always available, regardless of init stage or runtime stage.
## Solution B: Convert the warning in `get_current_vllm_config()` to a hard failure
But this means we may need to fix lots of CI failures.
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/30859 | open | [
"bug"
] | 2025-12-17T08:59:49Z | 2025-12-22T18:09:55Z | 7 | nvpohanh |
sgl-project/sglang | 15,319 | [Feature] RFC: AutoSpec, Automatic Runtime Speculative Inference Parameter Tuning | ### Checklist
- [x] If this is not a feature request but a general question, please start a discussion at https://github.com/sgl-project/sglang/discussions. Otherwise, it will be closed.
- [x] Please use English. Otherwise, it will be closed.
### Motivation
## Summary
This proposal introduces automatic runtime tuning for speculative inference parameters in SGLang. Instead of requiring users to manually set speculative_num_steps, speculative_topk, and speculative_num_draft_tokens, the system dynamically adjusts them using a feedback-driven controller. This maximizes throughput while respecting hardware limits and draft model capabilities—without any manual configuration.
## Problem & Motivation
Currently, users of speculative inference in SGLang must manually tune several parameters:
- speculative_num_steps
- speculative_topk
- speculative_num_draft_tokens
<img width="377" height="249" alt="Image" src="https://github.com/user-attachments/assets/d13c9e61-e5be-4d9c-a245-d9b95cb299f6" />
From the graph, we see that throughput varies with speculative_num_steps and batchsize, and it also suggests that a well-tuned parameter configuration of speculative inference can increase throughput by 5%~50%. These findings suggest three current issues:
1. Trial-and-error overhead – Finding optimal values per model/hardware/workload is tedious and often results in suboptimal performance.
2. Model capability mismatch – Different draft models have different effective limits, but static parameters cannot adapt.
3. Batch-size sensitivity – The optimal number of speculative steps decreases as batch size grows, due to compute constraints.
A single fixed configuration cannot perform well across varying models, hardware, and batch sizes.
## Proposed Design
We propose a lightweight feedback controller that adjusts speculative_num_steps in real time based on runtime metrics. For simplicity and stability, we keep speculative_topk=1 and speculative_num_draft_tokens=speculative_num_steps+1 (following observed best practices).
### Core Architecture
The system monitors two metrics after each batch:
- Acceptance rate – ratio of accepted draft tokens.
- Acceptance length growth – how much accepted length changes when steps increase.
Using these, it applies the following simple rules:
1. Increase steps if:
- Acceptance rate is high (configurable, e.g., ≥0.6)
- Acceptance length grows sufficiently (exceeding a model-aware threshold)
- Hardware limits for the current batch size are not exceeded
2. Decrease steps if:
- Acceptance rate is low (e.g., <0.5)
3. Otherwise, keep steps unchanged.
This forms a stable negative-feedback loop that converges to a near-optimal step count for the current workload.
### Detailed Designs
#### Initialization Phase
During system startup, the following initialization sequence occurs:
1. **Computational Threshold Calculation**: For each possible batch size (1, 2, 4, 8, 16, 32, 64), compute the maximum allowable speculative steps given hardware constraints(thres_batchsize);
2. **Draft Model Ability Analysis**: (Optional) Assess draft model capabilities and establish maximum effective step boundaries. (This step is optional, parameters can be dynamically adjusted and saved during runtime.)
3. **Theoretical Threshold Establishment**: Calculate lower bound of theoretical accept length growth thresholds for different speculative step values.
#### Runtime Parameter Adjustment Logic
The adjustment algorithm implements a conservative approach to prevent oscillation:
<img width="1384" height="1484" alt="Image" src="https://github.com/user-attachments/assets/f31880cc-7374-4818-9fa7-6aa6f4d1ed91" />
```
For each batch run:
1. Collect metrics: acceptance_rate, acceptance_length_growth_rate
2. speculative_num_steps += 1 if (acceptance_length_growth_rate > thres_accept_length_growth_rate AND accept_rate >= thres_positive_accept_rate AND speculative_num_step+1<thres_batchsize_num_steps)
3. speculative_num_steps -= 1 if (acceptance_length_growth_rate <= thres_accept_length_growth_rate OR accept_rate < thres_negative_accept_rate)
4. speculative_num_steps remains unchanged otherwise
5. Update parameters and record accept length of current loop
```
### Key Benefits
- Zero configuration – Users no longer need to guess parameters.
- Adaptive – Automatically adjusts to model pairs, hardware, and batch sizes.
- Performance-aware – Maximizes throughput while avoiding overload.
- Backward compatible – Manual configuration remains available.
## Command Line Arguments
| Argument | Description |
|----------|-------------|
| `--speculative-auto-tune` | Enable automatic tuning of speculative_num_steps (default: false) |
| `--speculative-min-steps` | Minimum speculative steps for dynamic adjustment (default: 1) |
| `--speculative-max-steps` | Maximum speculative steps for dynamic adjustment (default: 10) |
| `--speculative-positive-threshold` | Acceptance rate threshold for increasin | https://github.com/sgl-project/sglang/issues/15319 | open | [] | 2025-12-17T08:53:57Z | 2025-12-22T03:37:45Z | 3 | maodoudou168 |
vllm-project/vllm | 30,855 | [Usage]: Qwen3-30B-A3B-NVFP4 fails on Dell Pro Max GB10 with "no kernel image is available for execution on the device" | ### Your current environment
```
Hardware: Dell Pro Max GB10
OS: Ubuntu 24
CUDA: cuda_13.0.r13.0
Cuda compilation tools, release 13.0, V13.0.88;
vllm: V0.12.0
torch_version: 2.9.0+cu128
model: RedHatAI/Qwen3-30B-A3B-NVFP4 or nvidia/Qwen3-30B-A3B-NVFP4 or nvidia/Qwen3-30B-A3B-FP4
```
### How would you like to use vllm
### I'm trying to run the quantized model RedHatAI/Qwen3-30B-A3B-NVFP4 using vLLM v0.12.0 on a Dell Pro Max GB10.However, I get the following error during model loading: torch.AcceleratorError: CUDA error: no kernel image is available for execution on the device
vllm serve RedHatAI/Qwen3-30B-A3B-NVFP4 --port 8002 --gpu-memory-utilization 0.7
(APIServer pid=731925) INFO 12-17 16:03:13 [api_server.py:1772] vLLM API server version 0.12.0
(APIServer pid=731925) INFO 12-17 16:03:13 [utils.py:253] non-default args: {'model_tag': 'RedHatAI/Qwen3-30B-A3B-NVFP4', 'port': 8002, 'model': 'RedHatAI/Qwen3-30B-A3B-NVFP4', 'gpu_memory_utilization': 0.7}
(APIServer pid=731925) Downloading Model from https://www.modelscope.cn to directory: /home/smc01/.cache/modelscope/hub/models/RedHatAI/Qwen3-30B-A3B-NVFP4
(APIServer pid=731925) Downloading Model from https://www.modelscope.cn to directory: /home/smc01/.cache/modelscope/hub/models/RedHatAI/Qwen3-30B-A3B-NVFP4
(APIServer pid=731925) Downloading Model from https://www.modelscope.cn to directory: /home/smc01/.cache/modelscope/hub/models/RedHatAI/Qwen3-30B-A3B-NVFP4
(APIServer pid=731925) INFO 12-17 16:03:17 [model.py:637] Resolved architecture: Qwen3MoeForCausalLM
(APIServer pid=731925) INFO 12-17 16:03:17 [model.py:1750] Using max model len 40960
(APIServer pid=731925) INFO 12-17 16:03:17 [scheduler.py:228] Chunked prefill is enabled with max_num_batched_tokens=2048.
(APIServer pid=731925) Downloading Model from https://www.modelscope.cn to directory: /home/smc01/.cache/modelscope/hub/models/RedHatAI/Qwen3-30B-A3B-NVFP4
(APIServer pid=731925) Downloading Model from https://www.modelscope.cn to directory: /home/smc01/.cache/modelscope/hub/models/RedHatAI/Qwen3-30B-A3B-NVFP4
(EngineCore_DP0 pid=732093) INFO 12-17 16:03:22 [core.py:93] Initializing a V1 LLM engine (v0.12.0) with config: model='RedHatAI/Qwen3-30B-A3B-NVFP4', speculative_config=None, tokenizer='RedHatAI/Qwen3-30B-A3B-NVFP4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=40960, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, disable_custom_all_reduce=False, quantization=compressed-tensors, enforce_eager=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01), seed=0, served_model_name=RedHatAI/Qwen3-30B-A3B-NVFP4, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'level': None, 'mode': <CompilationMode.VLLM_COMPILE: 3>, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['none'], 'splitting_ops': ['vllm::unified_attention', 'vllm::unified_attention_with_output', 'vllm::unified_mla_attention', 'vllm::unified_mla_attention_with_output', 'vllm::mamba_mixer2', 'vllm::mamba_mixer', 'vllm::short_conv', 'vllm::linear_attention', 'vllm::plamo2_mamba_mixer', 'vllm::gdn_attention_core', 'vllm::kda_attention', 'vllm::sparse_attn_indexer'], 'compile_mm_encoder': False, 'compile_sizes': [], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': <CUDAGraphMode.FULL_AND_PIECEWISE: (2, 1)>, 'cudagraph_num_of_warmups': 1, 'cudagraph_capture_sizes': [1, 2, 4, 8, 16, 24, 32, 40, 48, 56, 64, 72, 80, 88, 96, 104, 112, 120, 128, 136, 144, 152, 160, 168, 176, 184, 192, 200, 208, 216, 224, 232, 240, 248, 256, 272, 288, 304, 320, 336, 352, 368, 384, 400, 416, 432, 448, 464, 480, 496, 512], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': False, 'fuse_act_quant': False, 'fuse_attn_quant': False, 'eliminate_noops': True, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False}, 'max_cudagraph_capture_size': 512, 'dynamic_shapes_config': {'type': <DynamicShapesType.BACKED: 'backed'>}, 'local_cache_dir': None}
(EngineCore_DP0 pid=732093) /home/smc01/miniconda3/envs/vLLM_12/lib/python3.10/site-packages/torch/cuda/__init__.py:283: UserWarning:
(EngineCore_DP0 pid=732093) Found GPU0 NVIDIA GB10 which is of cuda capabilit | https://github.com/vllm-project/vllm/issues/30855 | open | [
"usage"
] | 2025-12-17T08:44:11Z | 2025-12-17T08:44:11Z | 0 | nanbogong |
vllm-project/vllm | 30,847 | [Bug]: Qwen 3VL via Efficient Video Sampling (EVS) to trim video embeddings and found that the number of tokens after timestamp in the Prompt was not aligned with the actual number of tokens after pruning? | ### Your current environment
<details>
vllm serve Qwen3-VL-8B --video-pruning-rate=0.75
messages=[
{
"role": "user",
"content": [
# {"type": "text", "text": "What's in this video?"},
{"type": "text", "text": "这个视频和图片分别描述的是什么内容?"},
{
"type": "video_url",
"video_url": {
"url": "file:///codes/data/video/Tom_Jerry.mp4",
"fps": 1,
},
}
],
}
],
<summary>The output of <code>python collect_env.py</code></summary>
```text
The get-video_deplacement_qwen3vl method in the qwen3-vl.py file
Firstly: Calculate the number of frames per frame
Secondly, add the specific timestamp of<{cur_time:. 1f} s>to the Prompt and add the calculated number of tokens after the timestamp.
At this point, the number of tokens per frame is calculated based on the clipping rate, so except for the first frame, the number of tokens after each frame remains unchanged (EVS is not used to calculate the actual tokens here).
The EVS algorithm calculates that the number of tokens reserved for each frame is different. It will cause the number of tokens after timestamp to be inconsistent with the actual number of tokens after clipping
```
</details>
、
### 🐛 Describe the bug
1、get_video_replacement_qwen3vl
frames_idx_token=[165, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33, 33]
<img width="989" height="359" alt="Image" src="https://github.com/user-attachments/assets/aeb59c9e-b970-4a2c-851f-a001c8610f0a" />
2、compute_retention_mask
<img width="1058" height="160" alt="Image" src="https://github.com/user-attachments/assets/d96f2ce9-1ed2-4d5d-b5ac-28c6394d8ee0" />
<img width="1042" height="194" alt="Image" src="https://github.com/user-attachments/assets/2bc35d99-bbbb-4437-84fb-755567e97b40" />
3、embed_input_ids
<img width="988" height="421" alt="Image" src="https://github.com/user-attachments/assets/2f8fcd5b-9fef-49d5-b331-dd4f1f1d7335" />
input_ids:
<img width="886" height="624" alt="Image" src="https://github.com/user-attachments/assets/a62fcfe8-f29e-4e1d-a8fb-dec91c5e75e7" />
From the above 1 and 3, it can be seen that the data in frames_idx_token is the same as that in embed_input_ids,
The first frame contains 165 tokens, while the rest contain 33 tokens
151656 is the ID of the video token. The number of 151656 is the number of video tokens. The sum of the number of video tokens in all frames is the same as the sum of frames_idx_token.
Regarding the second item: compute_contention_mask EVS cropped mask, it was found that the number of tokens in the first frame was 165, while the number of tokens in other frames was different,
Based on the above 1, 2, and 3, it can be concluded that the current implementation of EVS pruning algorithm has problems
That is, the number of tokens after timestamp in the Prompt does not match the actual number of tokens that should be retained after EVS pruning.
### Before submitting a new issue...
- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions. | https://github.com/vllm-project/vllm/issues/30847 | open | [
"bug"
] | 2025-12-17T06:46:15Z | 2026-01-04T07:39:17Z | 5 | xshqhua |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.