vllm - ✅(Solved) Fix [Bug]: Validation Error: block_size (2096) > max_num_batched_tokens (2048) when enabling prefix caching for Qwen3.5 Mamba architecture [1 pull requests, 1 comments, 2 participants]
ON THIS PAGE
Recommended Tools
×6Utilities matched from this issue’s tags and category — try them while you read without losing context.
GitHub issue graph ai analysis
Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.
The report is written in English Markdown for sharing and archival.
Helpful · Quick feedback
GitHub stats
vllm-project/vllm#36697•Fetched 2026-04-08 00:35:20
Comments
1
Participants
2
Timeline
7
Reactions
0
Author
Imagium719Participants
flutistTimeline (top)
subscribed ×3commented ×1cross-referenced ×1labeled ×1
Error Message
(vllm) Yifei@lyg1058:~/models$ CUDA_VISIBLE_DEVICES=2,3 vllm serve Qwen3.5-122B-A10B-GPTQ-Int4 --served-model-name Qwen3.5-27B-AWQ-4bit --gpu-memory-utilization 0.9 --port 8848 -tp 2 --max-model-len 131072 --enable-auto-tool-choice --tool-call-parser qwen3_coder --reasoning-parser qwen3 --max-num-seqs 16 --enable-prefix-caching
(APIServer pid=913608) INFO 03-11 03:44:50 [utils.py:302]
(APIServer pid=913608) INFO 03-11 03:44:50 [utils.py:302] █ █ █▄ ▄█
(APIServer pid=913608) INFO 03-11 03:44:50 [utils.py:302] ▄▄ ▄█ █ █ █ ▀▄▀ █ version 0.17.0
(APIServer pid=913608) INFO 03-11 03:44:50 [utils.py:302] █▄█▀ █ █ █ █ model Qwen3.5-122B-A10B-GPTQ-Int4
(APIServer pid=913608) INFO 03-11 03:44:50 [utils.py:302] ▀▀ ▀▀▀▀▀ ▀▀▀▀▀ ▀ ▀
(APIServer pid=913608) INFO 03-11 03:44:50 [utils.py:302]
(APIServer pid=913608) INFO 03-11 03:44:50 [utils.py:238] non-default args: {'model_tag': 'Qwen3.5-122B-A10B-GPTQ-Int4', 'enable_auto_tool_choice': True, 'tool_call_parser': 'qwen3_coder', 'port': 8848, 'model': 'Qwen3.5-122B-A10B-GPTQ-Int4', 'max_model_len': 131072, 'served_model_name': ['Qwen3.5-27B-AWQ-4bit'], 'reasoning_parser': 'qwen3', 'tensor_parallel_size': 2, 'enable_prefix_caching': True, 'max_num_seqs': 16}
(APIServer pid=913608) INFO 03-11 03:44:59 [model.py:531] Resolved architecture: Qwen3_5MoeForConditionalGeneration
(APIServer pid=913608) INFO 03-11 03:44:59 [model.py:1554] Using max model len 131072
(APIServer pid=913608) INFO 03-11 03:44:59 [gptq_marlin.py:229] The model is convertible to gptq_marlin during runtime. Using gptq_marlin kernel.
(APIServer pid=913608) INFO 03-11 03:45:00 [scheduler.py:231] Chunked prefill is enabled with max_num_batched_tokens=2048.
(APIServer pid=913608) WARNING 03-11 03:45:00 [config.py:381] Mamba cache mode is set to 'align' for Qwen3_5MoeForConditionalGeneration by default when prefix caching is enabled
(APIServer pid=913608) INFO 03-11 03:45:00 [config.py:401] Warning: Prefix caching in Mamba cache 'align' mode is currently enabled. Its support for Mamba layers is experimental. Please report any issues you may observe.
(APIServer pid=913608) INFO 03-11 03:45:00 [config.py:544] Setting attention block size to 2096 tokens to ensure that attention page size is >= mamba page size.
(APIServer pid=913608) INFO 03-11 03:45:00 [config.py:575] Padding mamba page size by 0.58% to ensure that mamba page size and attention page size are exactly equal.
(APIServer pid=913608) INFO 03-11 03:45:01 [vllm.py:747] Asynchronous scheduling is enabled.
(APIServer pid=913608) Traceback (most recent call last):
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/bin/vllm", line 10, in <module>
(APIServer pid=913608) sys.exit(main())
(APIServer pid=913608) ^^^^^^
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/vllm/entrypoints/cli/main.py", line 73, in main
(APIServer pid=913608) args.dispatch_function(args)
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/vllm/entrypoints/cli/serve.py", line 112, in cmd
(APIServer pid=913608) uvloop.run(run_server(args))
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/uvloop/init.py", line 92, in run
(APIServer pid=913608) return runner.run(wrapper())
(APIServer pid=913608) ^^^^^^^^^^^^^^^^^^^^^
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/asyncio/runners.py", line 118, in run
(APIServer pid=913608) return self._loop.run_until_complete(task)
(APIServer pid=913608) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(APIServer pid=913608) File "uvloop/loop.pyx", line 1518, in uvloop.loop.Loop.run_until_complete
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/uvloop/init.py", line 48, in wrapper
(APIServer pid=913608) return await main
(APIServer pid=913608) ^^^^^^^^^^
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/vllm/entrypoints/openai/api_server.py", line 471, in run_server
(APIServer pid=913608) await run_server_worker(listen_address, sock, args, **uvicorn_kwargs)
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/vllm/entrypoints/openai/api_server.py", line 490, in run_server_worker
(APIServer pid=913608) async with build_async_engine_client(
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/contextlib.py", line 210, in aenter
(APIServer pid=913608) return await anext(self.gen)
(APIServer pid=913608) ^^^^^^^^^^^^^^^^^^^^^
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/vllm/entrypoints/openai/api_server.py", line 96, in build_async_engine_client
(APIServer pid=913608) async with build_async_engine_client_from_engine_args(
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/contextlib.py", line 210, in aenter
(APIServer pid=913608) return await anext(self.gen)
(APIServer pid=913608) ^^^^^^^^^^^^^^^^^^^^^
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/vllm/entrypoints/openai/api_server.py", line 122, in build_async_engine_client_from_engine_args
(APIServer pid=913608) vllm_config = engine_args.create_engine_config(usage_context=usage_context)
(APIServer pid=913608) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/vllm/engine/arg_utils.py", line 1890, in create_engine_config
(APIServer pid=913608) config = VllmConfig(
(APIServer pid=913608) ^^^^^^^^^^^
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/pydantic/_internal/_dataclasses.py", line 121, in init
(APIServer pid=913608) s.pydantic_validator.validate_python(ArgsKwargs(args, kwargs), self_instance=s)
(APIServer pid=913608) pydantic_core._pydantic_core.ValidationError: 1 validation error for VllmConfig
(APIServer pid=913608) Assertion failed, In Mamba cache align mode, block_size (2096) must be <= max_num_batched_tokens (2048). [type=assertion_error, input_value=ArgsKwargs((), {'model_co...transfer_config': None}), input_type=ArgsKwargs]
(APIServer pid=913608) For further information visit https://errors.pydantic.dev/2.12/v/assertion_error
Root Cause
- Bug summary: When running a Mamba-hybrid model (Qwen3_5MoeForConditionalGeneration) with --enable-prefix-caching, vLLM auto-scales the block_size to 2096 to align Mamba and Attention pages. However, it crashes during VllmConfig validation because the default max_num_batched_tokens for chunked prefill is 2048.
Fix Action
Fix / Workaround
============================== 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): 192 On-line CPU(s) list: 0-191 Vendor ID: GenuineIntel Model name: Intel(R) Xeon(R) Platinum 8468 CPU family: 6 Model: 143 Thread(s) per core: 2 Core(s) per socket: 48 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: 4.5 MiB (96 instances) L1i cache: 3 MiB (96 instances) L2 cache: 192 MiB (96 instances) L3 cache: 210 MiB (2 instances) NUMA node(s): 4 NUMA node0 CPU(s): 0-23,96-119 NUMA node1 CPU(s): 24-47,120-143 NUMA node2 CPU(s): 48-71,144-167 NUMA node3 CPU(s): 72-95,168-191 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 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 IBRS, IBPB conditional, RSB filling, PBRSB-eIBRS SW sequence Vulnerability Srbds: Not affected Vulnerability Tsx async abort: Not affected
(APIServer pid=913608) args.dispatch_function(args)
PR fix notes
PR #36734: Automatically increased max_num_batched_tokens under Mamba align mode
- Repository: vllm-project/vllm
- Author: flutist
- State: open | merged: False
- Link: https://github.com/vllm-project/vllm/pull/36734
Description (problem / solution / changelog)
Automatically increased max_num_batched_tokens to accommodate Mamba align mode block_size solve https://github.com/vllm-project/vllm/issues/36697 When using Mamba cache in align mode, block_size may exceed max_num_batched_tokens, causing alignment issues. This PR automatically bumps max_num_batched_tokens to match block_size when this condition is detected, and emits a warning to notify the user of the change.
<!-- markdownlint-disable -->Purpose
solve
Assertion failed, In Mamba cache align mode, block_size (2096) must be <= max_num_batched_tokens (2048). [type=assertion_error, input_value=ArgsKwargs((), {'model_co...transfer_config': None}), input_type=ArgsKwargs]
Test Plan
before hot fix, vllm serve Qwen/Qwen3.5-122B-A10B-GPTQ-Int4 --served-model-name Qwen3.5-27B-AWQ-4bit --gpu-memory-utilization 0.9 --port 8848 -tp 2 --max-model-len 131072 --enable-auto-tool-choice --tool-call-parser qwen3_coder --reasoning-parser qwen3 --max-num-seqs 16 --enable-prefix-caching show error
vllm serve Qwen/Qwen3.5-122B-A10B-GPTQ-Int4 --served-model-name Qwen3.5-27B-AWQ-4bit --gpu-memory-utilization 0.9 --port 8848 -tp 2 --max-model-len 131072 --enable-auto-tool-choice --tool-call-parser qwen3_coder --reasoning-parser qwen3 --max-num-seqs 16 --enable-prefix-caching
(APIServer pid=140954) INFO 03-17 11:29:48 [utils.py:302]
(APIServer pid=140954) INFO 03-17 11:29:48 [utils.py:302] █ █ █▄ ▄█
(APIServer pid=140954) INFO 03-17 11:29:48 [utils.py:302] ▄▄ ▄█ █ █ █ ▀▄▀ █ version 0.17.1
(APIServer pid=140954) INFO 03-17 11:29:48 [utils.py:302] █▄█▀ █ █ █ █ model Qwen/Qwen3.5-122B-A10B-GPTQ-Int4
(APIServer pid=140954) INFO 03-17 11:29:48 [utils.py:302] ▀▀ ▀▀▀▀▀ ▀▀▀▀▀ ▀ ▀
(APIServer pid=140954) INFO 03-17 11:29:48 [utils.py:302]
(APIServer pid=140954) INFO 03-17 11:29:48 [utils.py:238] non-default args: {'model_tag': 'Qwen/Qwen3.5-122B-A10B-GPTQ-Int4', 'enable_auto_tool_choice': True, 'tool_call_parser': 'qwen3_coder', 'port': 8848, 'model': 'Qwen/Qwen3.5-122B-A10B-GPTQ-Int4', 'max_model_len': 131072, 'served_model_name': ['Qwen3.5-27B-AWQ-4bit'], 'reasoning_parser': 'qwen3', 'tensor_parallel_size': 2, 'enable_prefix_caching': True, 'max_num_seqs': 16}
config.json: 4.33kB [00:00, 24.6MB/s]
preprocessor_config.json: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 390/390 [00:00<00:00, 4.23MB/s]
(APIServer pid=140954) INFO 03-17 11:29:55 [model.py:531] Resolved architecture: Qwen3_5MoeForConditionalGeneration
(APIServer pid=140954) INFO 03-17 11:29:55 [model.py:1554] Using max model len 131072
(APIServer pid=140954) INFO 03-17 11:29:55 [gptq_marlin.py:229] The model is convertible to gptq_marlin during runtime. Using gptq_marlin kernel.
(APIServer pid=140954) INFO 03-17 11:29:55 [scheduler.py:231] Chunked prefill is enabled with max_num_batched_tokens=2048.
(APIServer pid=140954) WARNING 03-17 11:29:55 [config.py:381] Mamba cache mode is set to 'align' for Qwen3_5MoeForConditionalGeneration by default when prefix caching is enabled
(APIServer pid=140954) INFO 03-17 11:29:55 [config.py:401] Warning: Prefix caching in Mamba cache 'align' mode is currently enabled. Its support for Mamba layers is experimental. Please report any issues you may observe.
(APIServer pid=140954) INFO 03-17 11:29:56 [config.py:544] Setting attention block size to 2096 tokens to ensure that attention page size is >= mamba page size.
(APIServer pid=140954) INFO 03-17 11:29:56 [config.py:575] Padding mamba page size by 0.58% to ensure that mamba page size and attention page size are exactly equal.
model.safetensors.index.json: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████| 18.1M/18.1M [00:02<00:00, 8.21MB/s]
Parse safetensors files: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 39/39 [00:02<00:00, 14.86it/s]
(APIServer pid=140954) INFO 03-17 11:30:02 [vllm.py:747] Asynchronous scheduling is enabled.
(APIServer pid=140954) Traceback (most recent call last):
(APIServer pid=140954) File "/home/admin/miniconda3/envs/for_test/bin/vllm", line 6, in <module>
(APIServer pid=140954) sys.exit(main())
(APIServer pid=140954) ^^^^^^
(APIServer pid=140954) File "/home/admin/miniconda3/envs/for_test/lib/python3.12/site-packages/vllm/entrypoints/cli/main.py", line 73, in main
(APIServer pid=140954) args.dispatch_function(args)
(APIServer pid=140954) File "/home/admin/miniconda3/envs/for_test/lib/python3.12/site-packages/vllm/entrypoints/cli/serve.py", line 112, in cmd
(APIServer pid=140954) uvloop.run(run_server(args))
(APIServer pid=140954) File "/home/admin/miniconda3/envs/for_test/lib/python3.12/site-packages/uvloop/__init__.py", line 96, in run
(APIServer pid=140954) return __asyncio.run(
(APIServer pid=140954) ^^^^^^^^^^^^^^
(APIServer pid=140954) File "/home/admin/miniconda3/envs/for_test/lib/python3.12/asyncio/runners.py", line 195, in run
(APIServer pid=140954) return runner.run(main)
(APIServer pid=140954) ^^^^^^^^^^^^^^^^
(APIServer pid=140954) File "/home/admin/miniconda3/envs/for_test/lib/python3.12/asyncio/runners.py", line 118, in run
(APIServer pid=140954) return self._loop.run_until_complete(task)
(APIServer pid=140954) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(APIServer pid=140954) File "uvloop/loop.pyx", line 1518, in uvloop.loop.Loop.run_until_complete
(APIServer pid=140954) File "/home/admin/miniconda3/envs/for_test/lib/python3.12/site-packages/uvloop/__init__.py", line 48, in wrapper
(APIServer pid=140954) return await main
(APIServer pid=140954) ^^^^^^^^^^
(APIServer pid=140954) File "/home/admin/miniconda3/envs/for_test/lib/python3.12/site-packages/vllm/entrypoints/openai/api_server.py", line 471, in run_server
(APIServer pid=140954) await run_server_worker(listen_address, sock, args, **uvicorn_kwargs)
(APIServer pid=140954) File "/home/admin/miniconda3/envs/for_test/lib/python3.12/site-packages/vllm/entrypoints/openai/api_server.py", line 490, in run_server_worker
(APIServer pid=140954) async with build_async_engine_client(
(APIServer pid=140954) ^^^^^^^^^^^^^^^^^^^^^^^^^^
(APIServer pid=140954) File "/home/admin/miniconda3/envs/for_test/lib/python3.12/contextlib.py", line 210, in __aenter__
(APIServer pid=140954) return await anext(self.gen)
(APIServer pid=140954) ^^^^^^^^^^^^^^^^^^^^^
(APIServer pid=140954) File "/home/admin/miniconda3/envs/for_test/lib/python3.12/site-packages/vllm/entrypoints/openai/api_server.py", line 96, in build_async_engine_client
(APIServer pid=140954) async with build_async_engine_client_from_engine_args(
(APIServer pid=140954) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(APIServer pid=140954) File "/home/admin/miniconda3/envs/for_test/lib/python3.12/contextlib.py", line 210, in __aenter__
(APIServer pid=140954) return await anext(self.gen)
(APIServer pid=140954) ^^^^^^^^^^^^^^^^^^^^^
(APIServer pid=140954) File "/home/admin/miniconda3/envs/for_test/lib/python3.12/site-packages/vllm/entrypoints/openai/api_server.py", line 122, in build_async_engine_client_from_engine_args
(APIServer pid=140954) vllm_config = engine_args.create_engine_config(usage_context=usage_context)
(APIServer pid=140954) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(APIServer pid=140954) File "/home/admin/miniconda3/envs/for_test/lib/python3.12/site-packages/vllm/engine/arg_utils.py", line 1890, in create_engine_config
(APIServer pid=140954) config = VllmConfig(
(APIServer pid=140954) ^^^^^^^^^^^
(APIServer pid=140954) File "/home/admin/miniconda3/envs/for_test/lib/python3.12/site-packages/pydantic/_internal/_dataclasses.py", line 121, in __init__
(APIServer pid=140954) s.__pydantic_validator__.validate_python(ArgsKwargs(args, kwargs), self_instance=s)
(APIServer pid=140954) pydantic_core._pydantic_core.ValidationError: 1 validation error for VllmConfig
(APIServer pid=140954) Assertion failed, In Mamba cache align mode, block_size (2096) must be <= max_num_batched_tokens (2048). [type=assertion_error, input_value=ArgsKwargs((), {'model_co...transfer_config': None}), input_type=ArgsKwargs]
(APIServer pid=140954) For further information visit https://errors.pydantic.dev/2.12/v/assertion_errorTest Result
After fix, it work.
vllm serve Qwen/Qwen3.5-122B-A10B-GPTQ-Int4 --served-model-name Qwen3.5-27B-AWQ-4bit --gpu-memory-utilization 0.9 --port 8848 -tp 2 --max-model-len 131072 --enable-auto-tool-choice --tool-call-parser qwen3_coder --reasoning-parser qwen3 --max-num-seqs 16 --enable-prefix-caching
(APIServer pid=147681) INFO 03-17 11:33:58 [utils.py:297]
(APIServer pid=147681) INFO 03-17 11:33:58 [utils.py:297] █ █ █▄ ▄█
(APIServer pid=147681) INFO 03-17 11:33:58 [utils.py:297] ▄▄ ▄█ █ █ █ ▀▄▀ █ version 0.1.dev14845+g57a314d15.d20260316
(APIServer pid=147681) INFO 03-17 11:33:58 [utils.py:297] █▄█▀ █ █ █ █ model Qwen/Qwen3.5-122B-A10B-GPTQ-Int4
(APIServer pid=147681) INFO 03-17 11:33:58 [utils.py:297] ▀▀ ▀▀▀▀▀ ▀▀▀▀▀ ▀ ▀
(APIServer pid=147681) INFO 03-17 11:33:58 [utils.py:297]
(APIServer pid=147681) INFO 03-17 11:33:58 [utils.py:233] non-default args: {'model_tag': 'Qwen/Qwen3.5-122B-A10B-GPTQ-Int4', 'enable_auto_tool_choice': True, 'tool_call_parser': 'qwen3_coder', 'port': 8848, 'model': 'Qwen/Qwen3.5-122B-A10B-GPTQ-Int4', 'max_model_len': 131072, 'served_model_name': ['Qwen3.5-27B-AWQ-4bit'], 'reasoning_parser': 'qwen3', 'tensor_parallel_size': 2, 'enable_prefix_caching': True, 'max_num_seqs': 16}
(APIServer pid=147681) INFO 03-17 11:34:10 [model.py:533] Resolved architecture: Qwen3_5MoeForConditionalGeneration
(APIServer pid=147681) INFO 03-17 11:34:10 [model.py:1582] Using max model len 131072
(APIServer pid=147681) INFO 03-17 11:34:12 [gptq_marlin.py:229] The model is convertible to gptq_marlin during runtime. Using gptq_marlin kernel.
(APIServer pid=147681) INFO 03-17 11:34:12 [scheduler.py:231] Chunked prefill is enabled with max_num_batched_tokens=2048.
(APIServer pid=147681) WARNING 03-17 11:34:12 [config.py:385] Mamba cache mode is set to 'align' for Qwen3_5MoeForConditionalGeneration by default when prefix caching is enabled
(APIServer pid=147681) INFO 03-17 11:34:12 [config.py:405] Warning: Prefix caching in Mamba cache 'align' mode is currently enabled. Its support for Mamba layers is experimental. Please report any issues you may observe.
(APIServer pid=147681) INFO 03-17 11:34:12 [config.py:212] Setting attention block size to 2096 tokens to ensure that attention page size is >= mamba page size.
(APIServer pid=147681) WARNING 03-17 11:34:12 [config.py:224] Automatically increased max_num_batched_tokens from 2048 to 2096 to accommodate Mamba align mode block_size
(APIServer pid=147681) INFO 03-17 11:34:12 [config.py:256] Padding mamba page size by 0.58% to ensure that mamba page size and attention page size are exactly equal.
Parse safetensors files: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 39/39 [00:01<00:00, 20.07it/s]
(APIServer pid=147681) INFO 03-17 11:34:16 [vllm.py:754] Asynchronous scheduling is enabled.
tokenizer_config.json: 16.7kB [00:00, 92.1MB/s]
vocab.json: 6.72MB [00:00, 143MB/s]
merges.txt: 3.35MB [00:00, 157MB/s]
tokenizer.json: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 12.8M/12.8M [00:01<00:00, 10.6MB/s]
chat_template.jinja: 7.76kB [00:00, 46.4MB/s]
generation_config.json: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 244/244 [00:00<00:00, 3.33MB/s]
video_preprocessor_config.json: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████| 385/385 [00:00<00:00, 4.40MB/s]
(EngineCore pid=149493) INFO 03-17 11:34:42 [core.py:103] Initializing a V1 LLM engine (v0.1.dev14845+g57a314d15.d20260316) with config: model='Qwen/Qwen3.5-122B-A10B-GPTQ-Int4', speculative_config=None, tokenizer='Qwen/Qwen3.5-122B-A10B-GPTQ-Int4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=131072, download_dir=None, load_format=auto, tensor_parallel_size=2, pipeline_parallel_size=1, data_parallel_size=1, decode_context_parallel_size=1, dcp_comm_backend=ag_rs, disable_custom_all_reduce=False, quantization=gptq_marlin, enforce_eager=False, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='qwen3', 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, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False), seed=0, served_model_name=Qwen3.5-27B-AWQ-4bit, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'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::olmo_hybrid_gdn_full_forward', 'vllm::kda_attention', 'vllm::sparse_attn_indexer', 'vllm::rocm_aiter_sparse_attn_indexer', 'vllm::unified_kv_cache_update', 'vllm::unified_mla_kv_cache_update'], 'compile_mm_encoder': False, 'compile_sizes': [], 'compile_ranges_endpoints': [2096], '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], '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, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False}, 'max_cudagraph_capture_size': 32, 'dynamic_shapes_config': {'type': <DynamicShapesType.BACKED: 'backed'>, 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': True, 'static_all_moe_layers': []}
(EngineCore pid=149493) WARNING 03-17 11:34:42 [multiproc_executor.py:997] Reducing Torch parallelism from 64 threads to 1 to avoid unnecessary CPU contention. Set OMP_NUM_THREADS in the external environment to tune this value as needed.
(EngineCore pid=149493) INFO 03-17 11:34:42 [multiproc_executor.py:134] DP group leader: node_rank=0, node_rank_within_dp=0, master_addr=127.0.0.1, mq_connect_ip=33.1.35.33 (local), world_size=2, local_world_size=2
(Worker pid=149761) INFO 03-17 11:34:52 [parallel_state.py:1395] world_size=2 rank=0 local_rank=0 distributed_init_method=tcp://127.0.0.1:47983 backend=nccl
(Worker pid=149883) INFO 03-17 11:34:55 [parallel_state.py:1395] world_size=2 rank=1 local_rank=1 distributed_init_method=tcp://127.0.0.1:47983 backend=nccl
(Worker pid=149883) <frozen importlib._bootstrap_external>:1301: FutureWarning: The cuda.cudart module is deprecated and will be removed in a future release, please switch to use the cuda.bindings.runtime module instead.
(Worker pid=149761) <frozen importlib._bootstrap_external>:1301: FutureWarning: The cuda.cudart module is deprecated and will be removed in a future release, please switch to use the cuda.bindings.runtime module instead.
(Worker pid=149883) <frozen importlib._bootstrap_external>:1301: FutureWarning: The cuda.nvrtc module is deprecated and will be removed in a future release, please switch to use the cuda.bindings.nvrtc module instead.
(Worker pid=149761) <frozen importlib._bootstrap_external>:1301: FutureWarning: The cuda.nvrtc module is deprecated and will be removed in a future release, please switch to use the cuda.bindings.nvrtc module instead.
(Worker pid=149761) INFO 03-17 11:34:56 [pynccl.py:111] vLLM is using nccl==2.27.5
(Worker pid=149761) WARNING 03-17 11:34:56 [symm_mem.py:67] SymmMemCommunicator: Device capability 8.9 not supported, communicator is not available.
(Worker pid=149883) WARNING 03-17 11:34:56 [symm_mem.py:67] SymmMemCommunicator: Device capability 8.9 not supported, communicator is not available.
(Worker pid=149883) INFO 03-17 11:34:56 [parallel_state.py:1717] rank 1 in world size 2 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 1, EP rank 1, EPLB rank N/A
(Worker pid=149761) INFO 03-17 11:34:56 [parallel_state.py:1717] rank 0 in world size 2 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank 0, EPLB rank N/A
(Worker_TP0 pid=149761) INFO 03-17 11:35:10 [gpu_model_runner.py:4492] Starting to load model Qwen/Qwen3.5-122B-A10B-GPTQ-Int4...
(Worker_TP1 pid=149883) INFO 03-17 11:35:10 [cuda.py:373] Using backend AttentionBackendEnum.FLASH_ATTN for vit attention
(Worker_TP1 pid=149883) INFO 03-17 11:35:10 [mm_encoder_attention.py:230] Using AttentionBackendEnum.FLASH_ATTN for MMEncoderAttention.
(Worker_TP0 pid=149761) INFO 03-17 11:35:10 [cuda.py:373] Using backend AttentionBackendEnum.FLASH_ATTN for vit attention
(Worker_TP0 pid=149761) INFO 03-17 11:35:10 [mm_encoder_attention.py:230] Using AttentionBackendEnum.FLASH_ATTN for MMEncoderAttention.
(Worker_TP1 pid=149883) INFO 03-17 11:35:11 [qwen3_next.py:198] Using Triton/FLA GDN prefill kernel
(Worker_TP0 pid=149761) INFO 03-17 11:35:11 [qwen3_next.py:198] Using Triton/FLA GDN prefill kernel
(Worker_TP0 pid=149761) INFO 03-17 11:35:11 [cuda.py:317] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION'].
(Worker_TP0 pid=149761) INFO 03-17 11:35:11 [flash_attn.py:593] Using FlashAttention version 2
model.safetensors-00005-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.69G/1.69G [00:20<00:00, 81.8MB/s]
model.safetensors-00003-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.69G/1.69G [00:20<00:00, 81.1MB/s]
model.safetensors-00004-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.69G/1.69G [00:20<00:00, 81.1MB/s]
model.safetensors-00002-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.69G/1.69G [00:21<00:00, 80.3MB/s]
model.safetensors-00007-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.69G/1.69G [00:22<00:00, 76.7MB/s]
model.safetensors-00006-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.69G/1.69G [00:26<00:00, 63.4MB/s]
model.safetensors-00001-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.69G/1.69G [00:27<00:00, 62.0MB/s]
model.safetensors-00008-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.69G/1.69G [00:27<00:00, 61.1MB/s]
model.safetensors-00009-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.69G/1.69G [00:19<00:00, 88.7MB/s]
model.safetensors-00013-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.69G/1.69G [00:18<00:00, 93.7MB/s]
model.safetensors-00010-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.69G/1.69G [00:21<00:00, 80.3MB/s]
model.safetensors-00012-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.69G/1.69G [00:22<00:00, 76.0MB/s]
model.safetensors-00011-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.69G/1.69G [00:23<00:00, 73.3MB/s]
model.safetensors-00014-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.69G/1.69G [00:17<00:00, 98.0MB/s]
model.safetensors-00015-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.69G/1.69G [00:18<00:00, 93.7MB/s]
model.safetensors-00016-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.69G/1.69G [00:24<00:00, 70.3MB/s]
model.safetensors-00017-of-00039.safeten(…): 100%|█████████████████████████████████████████████████████████████████████████████████████████████| 1.69G/1.69G [00:14<00:00, 115MB/s]
model.safetensors-00018-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.69G/1.69G [00:18<00:00, 90.7MB/s]
model.safetensors-00019-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.69G/1.69G [00:22<00:00, 76.6MB/s]
model.safetensors-00020-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.69G/1.69G [00:22<00:00, 74.6MB/s]
model.safetensors-00022-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.69G/1.69G [00:21<00:00, 77.3MB/s]
model.safetensors-00024-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.69G/1.69G [00:18<00:00, 93.7MB/s]
model.safetensors-00023-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.69G/1.69G [00:25<00:00, 67.5MB/s]
model.safetensors-00021-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.69G/1.69G [00:27<00:00, 60.7MB/s]
model.safetensors-00026-of-00039.safeten(…): 100%|█████████████████████████████████████████████████████████████████████████████████████████████| 1.68G/1.68G [00:12<00:00, 140MB/s]
model.safetensors-00025-of-00039.safeten(…): 100%|█████████████████████████████████████████████████████████████████████████████████████████████| 1.68G/1.68G [00:15<00:00, 106MB/s]
model.safetensors-00027-of-00039.safeten(…): 100%|█████████████████████████████████████████████████████████████████████████████████████████████| 1.68G/1.68G [00:13<00:00, 125MB/s]
model.safetensors-00029-of-00039.safeten(…): 100%|█████████████████████████████████████████████████████████████████████████████████████████████| 1.68G/1.68G [00:16<00:00, 105MB/s]
model.safetensors-00028-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.68G/1.68G [00:24<00:00, 68.2MB/s]
model.safetensors-00033-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.68G/1.68G [00:24<00:00, 69.9MB/s]
model.safetensors-00031-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.68G/1.68G [00:25<00:00, 67.1MB/s]
model.safetensors-00034-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.68G/1.68G [00:25<00:00, 65.0MB/s]
model.safetensors-00032-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.68G/1.68G [00:27<00:00, 61.2MB/s]
model.safetensors-00030-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.68G/1.68G [00:29<00:00, 57.1MB/s]
model.safetensors-00035-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.68G/1.68G [00:20<00:00, 80.7MB/s]
model.safetensors-00036-of-00039.safeten(…): 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1.68G/1.68G [00:23<00:00, 72.3MB/s]
model.safetensors-00037-of-00039.safeten(…): 100%|█████████████████████████████████████████████████████████████████████████████████████████████| 6.40G/6.40G [00:26<00:00, 242MB/s]
model.safetensors-00039-of-00039.safeten(…): 100%|█████████████████████████████████████████████████████████████████████████████████████████████| 5.41G/5.41G [00:31<00:00, 172MB/s]
model.safetensors-00038-of-00039.safeten(…): 100%|█████████████████████████████████████████████████████████████████████████████████████████████| 6.44G/6.44G [00:37<00:00, 174MB/s]
(Worker_TP1 pid=149883) INFO 03-17 11:37:49 [weight_utils.py:574] Time spent downloading weights for Qwen/Qwen3.5-122B-A10B-GPTQ-Int4: 152.912391 seconds40G [00:20<00:00, 295MB/s]
Loading safetensors checkpoint shards: 0% Completed | 0/39 [00:00<?, ?it/s]
Loading safetensors checkpoint shards: 3% Completed | 1/39 [00:13<08:14, 13.02s/it]
Loading safetensors checkpoint shards: 5% Completed | 2/39 [00:26<08:09, 13.22s/it]
Loading safetensors checkpoint shards: 8% Completed | 3/39 [00:37<07:27, 12.42s/it]
Loading safetensors checkpoint shards: 10% Completed | 4/39 [00:50<07:20, 12.60s/it]
Loading safetensors checkpoint shards: 13% Completed | 5/39 [01:04<07:27, 13.17s/it]
Loading safetensors checkpoint shards: 15% Completed | 6/39 [01:17<07:10, 13.06s/it]
Loading safetensors checkpoint shards: 18% Completed | 7/39 [01:30<06:58, 13.07s/it]
Loading safetensors checkpoint shards: 21% Completed | 8/39 [01:43<06:39, 12.89s/it]
Loading safetensors checkpoint shards: 23% Completed | 9/39 [01:54<06:12, 12.41s/it]
Loading safetensors checkpoint shards: 26% Completed | 10/39 [02:08<06:10, 12.79s/it]
Loading safetensors checkpoint shards: 28% Completed | 11/39 [02:22<06:07, 13.14s/it]
Loading safetensors checkpoint shards: 31% Completed | 12/39 [02:35<05:59, 13.32s/it]
Loading safetensors checkpoint shards: 33% Completed | 13/39 [02:50<05:56, 13.71s/it]
Loading safetensors checkpoint shards: 36% Completed | 14/39 [03:05<05:48, 13.95s/it]
Loading safetensors checkpoint shards: 38% Completed | 15/39 [03:19<05:36, 14.03s/it]
Loading safetensors checkpoint shards: 41% Completed | 16/39 [03:36<05:41, 14.86s/it]
Loading safetensors checkpoint shards: 44% Completed | 17/39 [03:51<05:33, 15.14s/it]
Loading safetensors checkpoint shards: 46% Completed | 18/39 [04:06<05:15, 15.04s/it]
Loading safetensors checkpoint shards: 49% Completed | 19/39 [04:18<04:42, 14.13s/it]
Loading safetensors checkpoint shards: 51% Completed | 20/39 [04:32<04:28, 14.11s/it]
Loading safetensors checkpoint shards: 54% Completed | 21/39 [04:46<04:11, 13.96s/it]
Loading safetensors checkpoint shards: 56% Completed | 22/39 [05:01<04:03, 14.33s/it]
Loading safetensors checkpoint shards: 59% Completed | 23/39 [05:16<03:50, 14.41s/it]
Loading safetensors checkpoint shards: 62% Completed | 24/39 [05:29<03:32, 14.19s/it]
Loading safetensors checkpoint shards: 64% Completed | 25/39 [05:42<03:10, 13.62s/it]
Loading safetensors checkpoint shards: 67% Completed | 26/39 [05:53<02:49, 13.04s/it]
Loading safetensors checkpoint shards: 69% Completed | 27/39 [06:03<02:25, 12.14s/it]
Loading safetensors checkpoint shards: 72% Completed | 28/39 [06:15<02:13, 12.09s/it]
Loading safetensors checkpoint shards: 74% Completed | 29/39 [06:30<02:08, 12.90s/it]
Loading safetensors checkpoint shards: 77% Completed | 30/39 [06:43<01:56, 12.90s/it]
Loading safetensors checkpoint shards: 79% Completed | 31/39 [06:57<01:45, 13.16s/it]
Loading safetensors checkpoint shards: 82% Completed | 32/39 [07:09<01:30, 12.89s/it]
Loading safetensors checkpoint shards: 85% Completed | 33/39 [07:21<01:15, 12.59s/it]
Loading safetensors checkpoint shards: 87% Completed | 34/39 [07:34<01:03, 12.67s/it]
Loading safetensors checkpoint shards: 90% Completed | 35/39 [07:52<00:57, 14.38s/it]
Loading safetensors checkpoint shards: 92% Completed | 36/39 [08:06<00:42, 14.31s/it]
Loading safetensors checkpoint shards: 95% Completed | 37/39 [08:45<00:43, 21.53s/it]
Loading safetensors checkpoint shards: 97% Completed | 38/39 [09:54<00:35, 35.91s/it]
Loading safetensors checkpoint shards: 100% Completed | 39/39 [10:12<00:00, 30.45s/it]
Loading safetensors checkpoint shards: 100% Completed | 39/39 [10:12<00:00, 15.70s/it]
(Worker_TP0 pid=149761)
(Worker_TP0 pid=149761) INFO 03-17 11:48:02 [default_loader.py:373] Loading weights took 612.40 seconds
(Worker_TP0 pid=149761) INFO 03-17 11:48:04 [gpu_model_runner.py:4577] Model loading took 34.32 GiB memory and 773.225700 seconds
(Worker_TP1 pid=149883) INFO 03-17 11:48:04 [gpu_model_runner.py:5499] Encoder cache will be initialized with a budget of 16384 tokens, and profiled with 1 image items of the maximum feature size.
(Worker_TP0 pid=149761) INFO 03-17 11:48:04 [gpu_model_runner.py:5499] Encoder cache will be initialized with a budget of 16384 tokens, and profiled with 1 image items of the maximum feature size.
(Worker_TP0 pid=149761) INFO 03-17 11:48:14 [backends.py:988] Using cache directory: /home/admin/.cache/vllm/torch_compile_cache/77532660c9/rank_0_0/backbone for vLLM's torch.compile
(Worker_TP0 pid=149761) INFO 03-17 11:48:14 [backends.py:1048] Dynamo bytecode transform time: 7.49 s
(Worker_TP1 pid=149883) INFO 03-17 11:48:18 [backends.py:371] Cache the graph of compile range (1, 2096) for later use
(Worker_TP0 pid=149761) INFO 03-17 11:48:18 [backends.py:371] Cache the graph of compile range (1, 2096) for later use
(Worker_TP0 pid=149761) INFO 03-17 11:48:37 [backends.py:387] Compiling a graph for compile range (1, 2096) takes 22.10 s
(Worker_TP1 pid=149883) /home/admin/workspace/aop_lab/app_source/vllm_custom_dataset_img_support_base64/vllm/model_executor/layers/fla/ops/utils.py:113: UserWarning: Input tensor shape suggests potential format mismatch: seq_len (16) < num_heads (32). This may indicate the inputs were passed in head-first format [B, H, T, ...] when head_first=False was specified. Please verify your input tensor format matches the expected shape [B, T, H, ...].
(Worker_TP1 pid=149883) return fn(*contiguous_args, **contiguous_kwargs)
(Worker_TP0 pid=149761) INFO 03-17 11:48:39 [decorators.py:627] saved AOT compiled function to /home/admin/.cache/vllm/torch_compile_cache/torch_aot_compile/6dcad07d427156b5cff90bf8a938953d4d4cc536dece8ca0f6cc17b416f33296/rank_0_0/model
(Worker_TP0 pid=149761) INFO 03-17 11:48:39 [monitor.py:48] torch.compile took 31.79 s in total
(Worker_TP0 pid=149761) /home/admin/workspace/aop_lab/app_source/vllm_custom_dataset_img_support_base64/vllm/model_executor/layers/fla/ops/utils.py:113: UserWarning: Input tensor shape suggests potential format mismatch: seq_len (16) < num_heads (32). This may indicate the inputs were passed in head-first format [B, H, T, ...] when head_first=False was specified. Please verify your input tensor format matches the expected shape [B, T, H, ...].
(Worker_TP0 pid=149761) return fn(*contiguous_args, **contiguous_kwargs)
(EngineCore pid=149493) INFO 03-17 11:49:05 [shm_broadcast.py:681] No available shared memory broadcast block found in 60 seconds. This typically happens when some processes are hanging or doing some time-consuming work (e.g. compilation, weight/kv cache quantization).
(Worker_TP0 pid=149761) /home/admin/workspace/aop_lab/app_source/vllm_custom_dataset_img_support_base64/vllm/model_executor/layers/fla/ops/utils.py:113: UserWarning: Input tensor shape suggests potential format mismatch: seq_len (16) < num_heads (32). This may indicate the inputs were passed in head-first format [B, H, T, ...] when head_first=False was specified. Please verify your input tensor format matches the expected shape [B, T, H, ...].
(Worker_TP0 pid=149761) return fn(*contiguous_args, **contiguous_kwargs)
(Worker_TP1 pid=149883) /home/admin/workspace/aop_lab/app_source/vllm_custom_dataset_img_support_base64/vllm/model_executor/layers/fla/ops/utils.py:113: UserWarning: Input tensor shape suggests potential format mismatch: seq_len (16) < num_heads (32). This may indicate the inputs were passed in head-first format [B, H, T, ...] when head_first=False was specified. Please verify your input tensor format matches the expected shape [B, T, H, ...].
(Worker_TP1 pid=149883) return fn(*contiguous_args, **contiguous_kwargs)
(Worker_TP0 pid=149761) INFO 03-17 11:50:04 [monitor.py:76] Initial profiling/warmup run took 85.05 s
(EngineCore pid=149493) INFO 03-17 11:50:05 [shm_broadcast.py:681] No available shared memory broadcast block found in 60 seconds. This typically happens when some processes are hanging or doing some time-consuming work (e.g. compilation, weight/kv cache quantization).
(Worker_TP1 pid=149883) INFO 03-17 11:50:11 [kv_cache_utils.py:826] Overriding num_gpu_blocks=0 with num_gpu_blocks_override=32
(Worker_TP0 pid=149761) INFO 03-17 11:50:11 [kv_cache_utils.py:826] Overriding num_gpu_blocks=0 with num_gpu_blocks_override=32
(Worker_TP1 pid=149883) INFO 03-17 11:50:11 [gpu_model_runner.py:5618] Profiling CUDA graph memory: PIECEWISE=7 (largest=32), FULL=5 (largest=16)
(Worker_TP0 pid=149761) INFO 03-17 11:50:11 [gpu_model_runner.py:5618] Profiling CUDA graph memory: PIECEWISE=7 (largest=32), FULL=5 (largest=16)
(Worker_TP0 pid=149761) INFO 03-17 11:50:13 [custom_all_reduce.py:216] Registering 384 cuda graph addresses
(Worker_TP1 pid=149883) INFO 03-17 11:50:13 [custom_all_reduce.py:216] Registering 384 cuda graph addresses
(Worker_TP0 pid=149761) INFO 03-17 11:50:14 [gpu_model_runner.py:5697] Estimated CUDA graph memory: 1.13 GiB total
(Worker_TP1 pid=149883) INFO 03-17 11:50:14 [gpu_model_runner.py:5697] Estimated CUDA graph memory: 1.13 GiB total
(Worker_TP1 pid=149883) INFO 03-17 11:50:15 [gpu_worker.py:486] In v0.19, CUDA graph memory profiling will be enabled by default (VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=1), which more accurately accounts for CUDA graph memory during KV cache allocation. To try it now, set VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=1 and increase --gpu-memory-utilization from 0.9000 to 0.9253 to maintain the same effective KV cache size.
(Worker_TP0 pid=149761) INFO 03-17 11:50:15 [gpu_worker.py:452] Available KV cache memory: 3.86 GiB
(Worker_TP0 pid=149761) INFO 03-17 11:50:15 [gpu_worker.py:486] In v0.19, CUDA graph memory profiling will be enabled by default (VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=1), which more accurately accounts for CUDA graph memory during KV cache allocation. To try it now, set VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=1 and increase --gpu-memory-utilization from 0.9000 to 0.9253 to maintain the same effective KV cache size.
(EngineCore pid=149493) INFO 03-17 11:50:15 [kv_cache_utils.py:1316] GPU KV cache size: 83,840 tokens
(EngineCore pid=149493) INFO 03-17 11:50:15 [kv_cache_utils.py:1321] Maximum concurrency for 131,072 tokens per request: 2.32x
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE): 100%|███████████████████████████████████████████████████████████████████████████████████████| 7/7 [00:01<00:00, 6.20it/s]
Capturing CUDA graphs (decode, FULL): 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████| 5/5 [00:00<00:00, 5.21it/s]
(Worker_TP1 pid=149883) INFO 03-17 11:50:17 [custom_all_reduce.py:216] Registering 1152 cuda graph addresses
(Worker_TP0 pid=149761) INFO 03-17 11:50:17 [custom_all_reduce.py:216] Registering 1152 cuda graph addresses
(Worker_TP0 pid=149761) INFO 03-17 11:50:18 [gpu_model_runner.py:5757] Graph capturing finished in 3 secs, took 0.78 GiB
(Worker_TP0 pid=149761) INFO 03-17 11:50:18 [gpu_worker.py:614] CUDA graph pool memory: 0.78 GiB (actual), 1.13 GiB (estimated), difference: 0.35 GiB (45.3%).
(Worker_TP1 pid=149883) INFO 03-17 11:50:18 [gpu_worker.py:614] CUDA graph pool memory: 0.78 GiB (actual), 1.13 GiB (estimated), difference: 0.35 GiB (45.3%).
(EngineCore pid=149493) INFO 03-17 11:50:18 [core.py:281] init engine (profile, create kv cache, warmup model) took 133.80 seconds
(EngineCore pid=149493) INFO 03-17 11:50:33 [vllm.py:754] Asynchronous scheduling is enabled.
(APIServer pid=147681) INFO 03-17 11:50:33 [api_server.py:573] Supported tasks: ['generate']
(APIServer pid=147681) INFO 03-17 11:50:35 [parser_manager.py:202] "auto" tool choice has been enabled.
(APIServer pid=147681) WARNING 03-17 11:50:35 [model.py:1376] Default vLLM sampling parameters have been overridden by the model's `generation_config.json`: `{'temperature': 0.6, 'top_k': 20, 'top_p': 0.95}`. If this is not intended, please relaunch vLLM instance with `--generation-config vllm`.
(APIServer pid=147681) INFO 03-17 11:50:35 [parser_manager.py:202] "auto" tool choice has been enabled.
(APIServer pid=147681) INFO 03-17 11:50:35 [parser_manager.py:202] "auto" tool choice has been enabled.
(APIServer pid=147681) INFO 03-17 11:50:41 [hf.py:320] Detected the chat template content format to be 'string'. You can set `--chat-template-content-format` to override this.
(APIServer pid=147681) INFO 03-17 11:50:49 [base.py:216] Multi-modal warmup completed in 7.982s
(APIServer pid=147681) INFO 03-17 11:50:49 [parser_manager.py:202] "auto" tool choice has been enabled.
(APIServer pid=147681) INFO 03-17 11:50:50 [api_server.py:577] Starting vLLM server on http://0.0.0.0:8848
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:37] Available routes are:
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /openapi.json, Methods: GET, HEAD
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /docs, Methods: GET, HEAD
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /docs/oauth2-redirect, Methods: GET, HEAD
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /redoc, Methods: GET, HEAD
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /tokenize, Methods: POST
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /detokenize, Methods: POST
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /load, Methods: GET
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /version, Methods: GET
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /health, Methods: GET
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /metrics, Methods: GET
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /v1/models, Methods: GET
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /ping, Methods: GET
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /ping, Methods: POST
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /invocations, Methods: POST
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /v1/chat/completions, Methods: POST
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /v1/responses, Methods: POST
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /v1/responses/{response_id}, Methods: GET
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /v1/responses/{response_id}/cancel, Methods: POST
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /v1/completions, Methods: POST
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /v1/messages, Methods: POST
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /v1/messages/count_tokens, Methods: POST
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /inference/v1/generate, Methods: POST
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /scale_elastic_ep, Methods: POST
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /is_scaling_elastic_ep, Methods: POST
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /v1/chat/completions/render, Methods: POST
(APIServer pid=147681) INFO 03-17 11:50:50 [launcher.py:46] Route: /v1/completions/render, Methods: POST
(APIServer pid=147681) INFO: Started server process [147681]
(APIServer pid=147681) INFO: Waiting for application startup.
(APIServer pid=147681) INFO: Application startup complete.And completion is work.
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8848/v1/chat/completions" \
> -H "Content-Type: application/json" \
> --data '{
> "model": "Qwen3.5-27B-AWQ-4bit",
> "messages": [
> {
> "role": "user",
> "content": "What is the capital of France?"
> }
> ]
> }'
{"id":"chatcmpl-9ad82c292b9872ab","object":"chat.completion","created":1773719595,"model":"Qwen3.5-27B-AWQ-4bit","choices":[{"index":0,"message":{"role":"assistant","content":"\n\nThe capital of France is **Paris**.","refusal":null,"annotations":null,"audio":null,"function_call":null,"tool_calls":[],"reasoning":"Thinking Process:\n\n1. **Identify the core question:** The user is asking \"What is the capital of France?\"\n2. **Retrieve knowledge:** Access general knowledge about geography and countries.\n3. **Verify the fact:** The capital of France is Paris.\n4. **Formulate the answer:** State the fact clearly and concisely.\n5. **Final Output:** \"The capital of France is Paris.\" or simply \"Paris.\" (A complete sentence is usually better for politeness).\n\n*Self-Correction/Refinement:* Keep it direct. No need for extra fluff unless asked.\n\n*Draft:* The capital of France is Paris.\n*Final Polish:* The capital of France is Paris.cw\n"},"logprobs":null,"finish_reason":"stop","stop_reason":null,"token_ids":null}],"service_tier":null,"system_fingerprint":null,"usage":{"prompt_tokens":17,"total_tokens":185,"completion_tokens":168,"prompt_tokens_details":null},"prompt_logprobs":null,"prompt_token_ids":null,"kv_transfer_params":null}(base)<details> <summary> Essential Elements of an Effective PR Description Checklist </summary>
- The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
- The test plan, such as providing test command.
- The test results, such as pasting the results comparison before and after, or e2e results
- (Optional) The necessary documentation update, such as updating
supported_models.mdandexamplesfor a new model. - (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.
Changed files
vllm/model_executor/models/config.py(modified, +13/-0)
Code Example
==============================
System Info
==============================
OS : Ubuntu 22.04.4 LTS (x86_64)
GCC version : (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version : Could not collect
CMake version : Could not collect
Libc version : glibc-2.35
==============================
PyTorch Info
==============================
PyTorch version : 2.10.0+cu128
Is debug build : False
CUDA used to build PyTorch : 12.8
ROCM used to build PyTorch : N/A
==============================
Python Environment
==============================
Python version : 3.11.14 (main, Oct 21 2025, 18:31:21) [GCC 11.2.0] (64-bit runtime)
Python platform : Linux-5.15.0-94-generic-x86_64-with-glibc2.35
==============================
CUDA / GPU Info
==============================
Is CUDA available : True
CUDA runtime version : 12.4.99
CUDA_MODULE_LOADING set to :
GPU models and configuration :
GPU 0: NVIDIA GeForce RTX 4090
GPU 1: NVIDIA GeForce RTX 4090
GPU 2: NVIDIA GeForce RTX 4090
GPU 3: NVIDIA GeForce RTX 4090
GPU 4: NVIDIA GeForce RTX 4090
GPU 5: NVIDIA GeForce RTX 4090
GPU 6: NVIDIA GeForce RTX 4090
GPU 7: NVIDIA GeForce RTX 4090
Nvidia driver version : 570.158.01
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): 192
On-line CPU(s) list: 0-191
Vendor ID: GenuineIntel
Model name: Intel(R) Xeon(R) Platinum 8468
CPU family: 6
Model: 143
Thread(s) per core: 2
Core(s) per socket: 48
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: 4.5 MiB (96 instances)
L1i cache: 3 MiB (96 instances)
L2 cache: 192 MiB (96 instances)
L3 cache: 210 MiB (2 instances)
NUMA node(s): 4
NUMA node0 CPU(s): 0-23,96-119
NUMA node1 CPU(s): 24-47,120-143
NUMA node2 CPU(s): 48-71,144-167
NUMA node3 CPU(s): 72-95,168-191
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 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 IBRS, IBPB conditional, RSB filling, PBRSB-eIBRS SW sequence
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
==============================
Versions of relevant libraries
==============================
[pip3] flashinfer-python==0.6.4
[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-cu12==12.8.90
[pip3] nvidia-cudnn-cu12==9.10.2.21
[pip3] nvidia-cudnn-frontend==1.19.0
[pip3] nvidia-cufft-cu12==11.3.3.83
[pip3] nvidia-cufile-cu12==1.13.1.3
[pip3] nvidia-curand-cu12==10.3.9.90
[pip3] nvidia-cusolver-cu12==11.7.3.90
[pip3] nvidia-cusparse-cu12==12.5.8.93
[pip3] nvidia-cusparselt-cu12==0.7.1
[pip3] nvidia-cutlass-dsl==4.4.1
[pip3] nvidia-cutlass-dsl-libs-base==4.4.1
[pip3] nvidia-ml-py==13.590.48
[pip3] nvidia-nccl-cu12==2.27.5
[pip3] nvidia-nvjitlink-cu12==12.8.93
[pip3] nvidia-nvshmem-cu12==3.4.5
[pip3] nvidia-nvtx-cu12==12.8.90
[pip3] pyzmq==27.1.0
[pip3] torch==2.10.0
[pip3] torch_c_dlpack_ext==0.1.5
[pip3] torchaudio==2.10.0
[pip3] torchvision==0.25.0
[pip3] transformers==4.57.6
[pip3] triton==3.6.0
[conda] flashinfer-python 0.6.4 pypi_0 pypi
[conda] numpy 2.2.6 pypi_0 pypi
[conda] nvidia-cublas-cu12 12.8.4.1 pypi_0 pypi
[conda] nvidia-cuda-cupti-cu12 12.8.90 pypi_0 pypi
[conda] nvidia-cuda-nvrtc-cu12 12.8.93 pypi_0 pypi
[conda] nvidia-cuda-runtime-cu12 12.8.90 pypi_0 pypi
[conda] nvidia-cudnn-cu12 9.10.2.21 pypi_0 pypi
[conda] nvidia-cudnn-frontend 1.19.0 pypi_0 pypi
[conda] nvidia-cufft-cu12 11.3.3.83 pypi_0 pypi
[conda] nvidia-cufile-cu12 1.13.1.3 pypi_0 pypi
[conda] nvidia-curand-cu12 10.3.9.90 pypi_0 pypi
[conda] nvidia-cusolver-cu12 11.7.3.90 pypi_0 pypi
[conda] nvidia-cusparse-cu12 12.5.8.93 pypi_0 pypi
[conda] nvidia-cusparselt-cu12 0.7.1 pypi_0 pypi
[conda] nvidia-cutlass-dsl 4.4.1 pypi_0 pypi
[conda] nvidia-cutlass-dsl-libs-base 4.4.1 pypi_0 pypi
[conda] nvidia-ml-py 13.590.48 pypi_0 pypi
[conda] nvidia-nccl-cu12 2.27.5 pypi_0 pypi
[conda] nvidia-nvjitlink-cu12 12.8.93 pypi_0 pypi
[conda] nvidia-nvshmem-cu12 3.4.5 pypi_0 pypi
[conda] nvidia-nvtx-cu12 12.8.90 pypi_0 pypi
[conda] pyzmq 27.1.0 pypi_0 pypi
[conda] torch 2.10.0 pypi_0 pypi
[conda] torch-c-dlpack-ext 0.1.5 pypi_0 pypi
[conda] torchaudio 2.10.0 pypi_0 pypi
[conda] torchvision 0.25.0 pypi_0 pypi
[conda] transformers 4.57.6 pypi_0 pypi
[conda] triton 3.6.0 pypi_0 pypi
==============================
vLLM Info
==============================
ROCM Version : Could not collect
vLLM Version : 0.17.0
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled
GPU Topology:
GPU0 GPU1 GPU2 GPU3 GPU4 GPU5 GPU6 GPU7 NIC0 CPU Affinity NUMA Affinity GPU NUMA ID
GPU0 X NODE NODE NODE SYS SYS SYS SYS SYS 0-23,96-119 0 N/A
GPU1 NODE X NODE NODE SYS SYS SYS SYS SYS 0-23,96-119 0 N/A
GPU2 NODE NODE X NODE SYS SYS SYS SYS SYS 0-23,96-119 0 N/A
GPU3 NODE NODE NODE X SYS SYS SYS SYS SYS 0-23,96-119 0 N/A
GPU4 SYS SYS SYS SYS X NODE NODE NODE NODE 24-47,120-143 1 N/A
GPU5 SYS SYS SYS SYS NODE X NODE NODE NODE 24-47,120-143 1 N/A
GPU6 SYS SYS SYS SYS NODE NODE X NODE NODE 24-47,120-143 1 N/A
GPU7 SYS SYS SYS SYS NODE NODE NODE X NODE 24-47,120-143 1 N/A
NIC0 SYS SYS SYS SYS NODE NODE NODE NODE 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
==============================
Environment Variables
==============================
LD_LIBRARY_PATH=/usr/local/cuda-12.4/lib64:
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
TORCHINDUCTOR_CACHE_DIR=/tmp/torchinductor_Yifei
---
(vllm) Yifei@lyg1058:~/models$ CUDA_VISIBLE_DEVICES=2,3 vllm serve Qwen3.5-122B-A10B-GPTQ-Int4 --served-model-name Qwen3.5-27B-AWQ-4bit --gpu-memory-utilization 0.9 --port 8848 -tp 2 --max-model-len 131072 --enable-auto-tool-choice --tool-call-parser qwen3_coder --reasoning-parser qwen3 --max-num-seqs 16 --enable-prefix-caching
(APIServer pid=913608) INFO 03-11 03:44:50 [utils.py:302]
(APIServer pid=913608) INFO 03-11 03:44:50 [utils.py:302] █ █ █▄ ▄█
(APIServer pid=913608) INFO 03-11 03:44:50 [utils.py:302] ▄▄ ▄█ █ █ █ ▀▄▀ █ version 0.17.0
(APIServer pid=913608) INFO 03-11 03:44:50 [utils.py:302] █▄█▀ █ █ █ █ model Qwen3.5-122B-A10B-GPTQ-Int4
(APIServer pid=913608) INFO 03-11 03:44:50 [utils.py:302] ▀▀ ▀▀▀▀▀ ▀▀▀▀▀ ▀ ▀
(APIServer pid=913608) INFO 03-11 03:44:50 [utils.py:302]
(APIServer pid=913608) INFO 03-11 03:44:50 [utils.py:238] non-default args: {'model_tag': 'Qwen3.5-122B-A10B-GPTQ-Int4', 'enable_auto_tool_choice': True, 'tool_call_parser': 'qwen3_coder', 'port': 8848, 'model': 'Qwen3.5-122B-A10B-GPTQ-Int4', 'max_model_len': 131072, 'served_model_name': ['Qwen3.5-27B-AWQ-4bit'], 'reasoning_parser': 'qwen3', 'tensor_parallel_size': 2, 'enable_prefix_caching': True, 'max_num_seqs': 16}
(APIServer pid=913608) INFO 03-11 03:44:59 [model.py:531] Resolved architecture: Qwen3_5MoeForConditionalGeneration
(APIServer pid=913608) INFO 03-11 03:44:59 [model.py:1554] Using max model len 131072
(APIServer pid=913608) INFO 03-11 03:44:59 [gptq_marlin.py:229] The model is convertible to gptq_marlin during runtime. Using gptq_marlin kernel.
(APIServer pid=913608) INFO 03-11 03:45:00 [scheduler.py:231] Chunked prefill is enabled with max_num_batched_tokens=2048.
(APIServer pid=913608) WARNING 03-11 03:45:00 [config.py:381] Mamba cache mode is set to 'align' for Qwen3_5MoeForConditionalGeneration by default when prefix caching is enabled
(APIServer pid=913608) INFO 03-11 03:45:00 [config.py:401] Warning: Prefix caching in Mamba cache 'align' mode is currently enabled. Its support for Mamba layers is experimental. Please report any issues you may observe.
(APIServer pid=913608) INFO 03-11 03:45:00 [config.py:544] Setting attention block size to 2096 tokens to ensure that attention page size is >= mamba page size.
(APIServer pid=913608) INFO 03-11 03:45:00 [config.py:575] Padding mamba page size by 0.58% to ensure that mamba page size and attention page size are exactly equal.
(APIServer pid=913608) INFO 03-11 03:45:01 [vllm.py:747] Asynchronous scheduling is enabled.
(APIServer pid=913608) Traceback (most recent call last):
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/bin/vllm", line 10, in <module>
(APIServer pid=913608) sys.exit(main())
(APIServer pid=913608) ^^^^^^
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/vllm/entrypoints/cli/main.py", line 73, in main
(APIServer pid=913608) args.dispatch_function(args)
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/vllm/entrypoints/cli/serve.py", line 112, in cmd
(APIServer pid=913608) uvloop.run(run_server(args))
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/uvloop/__init__.py", line 92, in run
(APIServer pid=913608) return runner.run(wrapper())
(APIServer pid=913608) ^^^^^^^^^^^^^^^^^^^^^
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/asyncio/runners.py", line 118, in run
(APIServer pid=913608) return self._loop.run_until_complete(task)
(APIServer pid=913608) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(APIServer pid=913608) File "uvloop/loop.pyx", line 1518, in uvloop.loop.Loop.run_until_complete
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/uvloop/__init__.py", line 48, in wrapper
(APIServer pid=913608) return await main
(APIServer pid=913608) ^^^^^^^^^^
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/vllm/entrypoints/openai/api_server.py", line 471, in run_server
(APIServer pid=913608) await run_server_worker(listen_address, sock, args, **uvicorn_kwargs)
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/vllm/entrypoints/openai/api_server.py", line 490, in run_server_worker
(APIServer pid=913608) async with build_async_engine_client(
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/contextlib.py", line 210, in __aenter__
(APIServer pid=913608) return await anext(self.gen)
(APIServer pid=913608) ^^^^^^^^^^^^^^^^^^^^^
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/vllm/entrypoints/openai/api_server.py", line 96, in build_async_engine_client
(APIServer pid=913608) async with build_async_engine_client_from_engine_args(
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/contextlib.py", line 210, in __aenter__
(APIServer pid=913608) return await anext(self.gen)
(APIServer pid=913608) ^^^^^^^^^^^^^^^^^^^^^
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/vllm/entrypoints/openai/api_server.py", line 122, in build_async_engine_client_from_engine_args
(APIServer pid=913608) vllm_config = engine_args.create_engine_config(usage_context=usage_context)
(APIServer pid=913608) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/vllm/engine/arg_utils.py", line 1890, in create_engine_config
(APIServer pid=913608) config = VllmConfig(
(APIServer pid=913608) ^^^^^^^^^^^
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/pydantic/_internal/_dataclasses.py", line 121, in __init__
(APIServer pid=913608) s.__pydantic_validator__.validate_python(ArgsKwargs(args, kwargs), self_instance=s)
(APIServer pid=913608) pydantic_core._pydantic_core.ValidationError: 1 validation error for VllmConfig
(APIServer pid=913608) Assertion failed, In Mamba cache align mode, block_size (2096) must be <= max_num_batched_tokens (2048). [type=assertion_error, input_value=ArgsKwargs((), {'model_co...transfer_config': None}), input_type=ArgsKwargs]
(APIServer pid=913608) For further information visit https://errors.pydantic.dev/2.12/v/assertion_errorRAW_BUFFERClick to expand / collapse
Your current environment
<details> <summary>The output of <code>python collect_env.py</code></summary>==============================
System Info
==============================
OS : Ubuntu 22.04.4 LTS (x86_64)
GCC version : (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version : Could not collect
CMake version : Could not collect
Libc version : glibc-2.35
==============================
PyTorch Info
==============================
PyTorch version : 2.10.0+cu128
Is debug build : False
CUDA used to build PyTorch : 12.8
ROCM used to build PyTorch : N/A
==============================
Python Environment
==============================
Python version : 3.11.14 (main, Oct 21 2025, 18:31:21) [GCC 11.2.0] (64-bit runtime)
Python platform : Linux-5.15.0-94-generic-x86_64-with-glibc2.35
==============================
CUDA / GPU Info
==============================
Is CUDA available : True
CUDA runtime version : 12.4.99
CUDA_MODULE_LOADING set to :
GPU models and configuration :
GPU 0: NVIDIA GeForce RTX 4090
GPU 1: NVIDIA GeForce RTX 4090
GPU 2: NVIDIA GeForce RTX 4090
GPU 3: NVIDIA GeForce RTX 4090
GPU 4: NVIDIA GeForce RTX 4090
GPU 5: NVIDIA GeForce RTX 4090
GPU 6: NVIDIA GeForce RTX 4090
GPU 7: NVIDIA GeForce RTX 4090
Nvidia driver version : 570.158.01
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): 192
On-line CPU(s) list: 0-191
Vendor ID: GenuineIntel
Model name: Intel(R) Xeon(R) Platinum 8468
CPU family: 6
Model: 143
Thread(s) per core: 2
Core(s) per socket: 48
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: 4.5 MiB (96 instances)
L1i cache: 3 MiB (96 instances)
L2 cache: 192 MiB (96 instances)
L3 cache: 210 MiB (2 instances)
NUMA node(s): 4
NUMA node0 CPU(s): 0-23,96-119
NUMA node1 CPU(s): 24-47,120-143
NUMA node2 CPU(s): 48-71,144-167
NUMA node3 CPU(s): 72-95,168-191
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 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 IBRS, IBPB conditional, RSB filling, PBRSB-eIBRS SW sequence
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
==============================
Versions of relevant libraries
==============================
[pip3] flashinfer-python==0.6.4
[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-cu12==12.8.90
[pip3] nvidia-cudnn-cu12==9.10.2.21
[pip3] nvidia-cudnn-frontend==1.19.0
[pip3] nvidia-cufft-cu12==11.3.3.83
[pip3] nvidia-cufile-cu12==1.13.1.3
[pip3] nvidia-curand-cu12==10.3.9.90
[pip3] nvidia-cusolver-cu12==11.7.3.90
[pip3] nvidia-cusparse-cu12==12.5.8.93
[pip3] nvidia-cusparselt-cu12==0.7.1
[pip3] nvidia-cutlass-dsl==4.4.1
[pip3] nvidia-cutlass-dsl-libs-base==4.4.1
[pip3] nvidia-ml-py==13.590.48
[pip3] nvidia-nccl-cu12==2.27.5
[pip3] nvidia-nvjitlink-cu12==12.8.93
[pip3] nvidia-nvshmem-cu12==3.4.5
[pip3] nvidia-nvtx-cu12==12.8.90
[pip3] pyzmq==27.1.0
[pip3] torch==2.10.0
[pip3] torch_c_dlpack_ext==0.1.5
[pip3] torchaudio==2.10.0
[pip3] torchvision==0.25.0
[pip3] transformers==4.57.6
[pip3] triton==3.6.0
[conda] flashinfer-python 0.6.4 pypi_0 pypi
[conda] numpy 2.2.6 pypi_0 pypi
[conda] nvidia-cublas-cu12 12.8.4.1 pypi_0 pypi
[conda] nvidia-cuda-cupti-cu12 12.8.90 pypi_0 pypi
[conda] nvidia-cuda-nvrtc-cu12 12.8.93 pypi_0 pypi
[conda] nvidia-cuda-runtime-cu12 12.8.90 pypi_0 pypi
[conda] nvidia-cudnn-cu12 9.10.2.21 pypi_0 pypi
[conda] nvidia-cudnn-frontend 1.19.0 pypi_0 pypi
[conda] nvidia-cufft-cu12 11.3.3.83 pypi_0 pypi
[conda] nvidia-cufile-cu12 1.13.1.3 pypi_0 pypi
[conda] nvidia-curand-cu12 10.3.9.90 pypi_0 pypi
[conda] nvidia-cusolver-cu12 11.7.3.90 pypi_0 pypi
[conda] nvidia-cusparse-cu12 12.5.8.93 pypi_0 pypi
[conda] nvidia-cusparselt-cu12 0.7.1 pypi_0 pypi
[conda] nvidia-cutlass-dsl 4.4.1 pypi_0 pypi
[conda] nvidia-cutlass-dsl-libs-base 4.4.1 pypi_0 pypi
[conda] nvidia-ml-py 13.590.48 pypi_0 pypi
[conda] nvidia-nccl-cu12 2.27.5 pypi_0 pypi
[conda] nvidia-nvjitlink-cu12 12.8.93 pypi_0 pypi
[conda] nvidia-nvshmem-cu12 3.4.5 pypi_0 pypi
[conda] nvidia-nvtx-cu12 12.8.90 pypi_0 pypi
[conda] pyzmq 27.1.0 pypi_0 pypi
[conda] torch 2.10.0 pypi_0 pypi
[conda] torch-c-dlpack-ext 0.1.5 pypi_0 pypi
[conda] torchaudio 2.10.0 pypi_0 pypi
[conda] torchvision 0.25.0 pypi_0 pypi
[conda] transformers 4.57.6 pypi_0 pypi
[conda] triton 3.6.0 pypi_0 pypi
==============================
vLLM Info
==============================
ROCM Version : Could not collect
vLLM Version : 0.17.0
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled
GPU Topology:
GPU0 GPU1 GPU2 GPU3 GPU4 GPU5 GPU6 GPU7 NIC0 CPU Affinity NUMA Affinity GPU NUMA ID
GPU0 X NODE NODE NODE SYS SYS SYS SYS SYS 0-23,96-119 0 N/A
GPU1 NODE X NODE NODE SYS SYS SYS SYS SYS 0-23,96-119 0 N/A
GPU2 NODE NODE X NODE SYS SYS SYS SYS SYS 0-23,96-119 0 N/A
GPU3 NODE NODE NODE X SYS SYS SYS SYS SYS 0-23,96-119 0 N/A
GPU4 SYS SYS SYS SYS X NODE NODE NODE NODE 24-47,120-143 1 N/A
GPU5 SYS SYS SYS SYS NODE X NODE NODE NODE 24-47,120-143 1 N/A
GPU6 SYS SYS SYS SYS NODE NODE X NODE NODE 24-47,120-143 1 N/A
GPU7 SYS SYS SYS SYS NODE NODE NODE X NODE 24-47,120-143 1 N/A
NIC0 SYS SYS SYS SYS NODE NODE NODE NODE 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
==============================
Environment Variables
==============================
LD_LIBRARY_PATH=/usr/local/cuda-12.4/lib64:
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
TORCHINDUCTOR_CACHE_DIR=/tmp/torchinductor_Yifei🐛 Describe the bug
-
Bug summary: When running a Mamba-hybrid model (Qwen3_5MoeForConditionalGeneration) with --enable-prefix-caching, vLLM auto-scales the block_size to 2096 to align Mamba and Attention pages. However, it crashes during VllmConfig validation because the default max_num_batched_tokens for chunked prefill is 2048.
-
Suggested fix: The framework should either automatically scale max_num_batched_tokens to be at least block_size when Mamba align mode forces it, or provide a clearer warning to the user to manually increase --max-num-batched-tokens.
-
Full Traceback:
(vllm) Yifei@lyg1058:~/models$ CUDA_VISIBLE_DEVICES=2,3 vllm serve Qwen3.5-122B-A10B-GPTQ-Int4 --served-model-name Qwen3.5-27B-AWQ-4bit --gpu-memory-utilization 0.9 --port 8848 -tp 2 --max-model-len 131072 --enable-auto-tool-choice --tool-call-parser qwen3_coder --reasoning-parser qwen3 --max-num-seqs 16 --enable-prefix-caching
(APIServer pid=913608) INFO 03-11 03:44:50 [utils.py:302]
(APIServer pid=913608) INFO 03-11 03:44:50 [utils.py:302] █ █ █▄ ▄█
(APIServer pid=913608) INFO 03-11 03:44:50 [utils.py:302] ▄▄ ▄█ █ █ █ ▀▄▀ █ version 0.17.0
(APIServer pid=913608) INFO 03-11 03:44:50 [utils.py:302] █▄█▀ █ █ █ █ model Qwen3.5-122B-A10B-GPTQ-Int4
(APIServer pid=913608) INFO 03-11 03:44:50 [utils.py:302] ▀▀ ▀▀▀▀▀ ▀▀▀▀▀ ▀ ▀
(APIServer pid=913608) INFO 03-11 03:44:50 [utils.py:302]
(APIServer pid=913608) INFO 03-11 03:44:50 [utils.py:238] non-default args: {'model_tag': 'Qwen3.5-122B-A10B-GPTQ-Int4', 'enable_auto_tool_choice': True, 'tool_call_parser': 'qwen3_coder', 'port': 8848, 'model': 'Qwen3.5-122B-A10B-GPTQ-Int4', 'max_model_len': 131072, 'served_model_name': ['Qwen3.5-27B-AWQ-4bit'], 'reasoning_parser': 'qwen3', 'tensor_parallel_size': 2, 'enable_prefix_caching': True, 'max_num_seqs': 16}
(APIServer pid=913608) INFO 03-11 03:44:59 [model.py:531] Resolved architecture: Qwen3_5MoeForConditionalGeneration
(APIServer pid=913608) INFO 03-11 03:44:59 [model.py:1554] Using max model len 131072
(APIServer pid=913608) INFO 03-11 03:44:59 [gptq_marlin.py:229] The model is convertible to gptq_marlin during runtime. Using gptq_marlin kernel.
(APIServer pid=913608) INFO 03-11 03:45:00 [scheduler.py:231] Chunked prefill is enabled with max_num_batched_tokens=2048.
(APIServer pid=913608) WARNING 03-11 03:45:00 [config.py:381] Mamba cache mode is set to 'align' for Qwen3_5MoeForConditionalGeneration by default when prefix caching is enabled
(APIServer pid=913608) INFO 03-11 03:45:00 [config.py:401] Warning: Prefix caching in Mamba cache 'align' mode is currently enabled. Its support for Mamba layers is experimental. Please report any issues you may observe.
(APIServer pid=913608) INFO 03-11 03:45:00 [config.py:544] Setting attention block size to 2096 tokens to ensure that attention page size is >= mamba page size.
(APIServer pid=913608) INFO 03-11 03:45:00 [config.py:575] Padding mamba page size by 0.58% to ensure that mamba page size and attention page size are exactly equal.
(APIServer pid=913608) INFO 03-11 03:45:01 [vllm.py:747] Asynchronous scheduling is enabled.
(APIServer pid=913608) Traceback (most recent call last):
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/bin/vllm", line 10, in <module>
(APIServer pid=913608) sys.exit(main())
(APIServer pid=913608) ^^^^^^
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/vllm/entrypoints/cli/main.py", line 73, in main
(APIServer pid=913608) args.dispatch_function(args)
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/vllm/entrypoints/cli/serve.py", line 112, in cmd
(APIServer pid=913608) uvloop.run(run_server(args))
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/uvloop/__init__.py", line 92, in run
(APIServer pid=913608) return runner.run(wrapper())
(APIServer pid=913608) ^^^^^^^^^^^^^^^^^^^^^
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/asyncio/runners.py", line 118, in run
(APIServer pid=913608) return self._loop.run_until_complete(task)
(APIServer pid=913608) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(APIServer pid=913608) File "uvloop/loop.pyx", line 1518, in uvloop.loop.Loop.run_until_complete
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/uvloop/__init__.py", line 48, in wrapper
(APIServer pid=913608) return await main
(APIServer pid=913608) ^^^^^^^^^^
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/vllm/entrypoints/openai/api_server.py", line 471, in run_server
(APIServer pid=913608) await run_server_worker(listen_address, sock, args, **uvicorn_kwargs)
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/vllm/entrypoints/openai/api_server.py", line 490, in run_server_worker
(APIServer pid=913608) async with build_async_engine_client(
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/contextlib.py", line 210, in __aenter__
(APIServer pid=913608) return await anext(self.gen)
(APIServer pid=913608) ^^^^^^^^^^^^^^^^^^^^^
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/vllm/entrypoints/openai/api_server.py", line 96, in build_async_engine_client
(APIServer pid=913608) async with build_async_engine_client_from_engine_args(
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/contextlib.py", line 210, in __aenter__
(APIServer pid=913608) return await anext(self.gen)
(APIServer pid=913608) ^^^^^^^^^^^^^^^^^^^^^
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/vllm/entrypoints/openai/api_server.py", line 122, in build_async_engine_client_from_engine_args
(APIServer pid=913608) vllm_config = engine_args.create_engine_config(usage_context=usage_context)
(APIServer pid=913608) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/vllm/engine/arg_utils.py", line 1890, in create_engine_config
(APIServer pid=913608) config = VllmConfig(
(APIServer pid=913608) ^^^^^^^^^^^
(APIServer pid=913608) File "/data/home/Yifei/miniconda3/envs/vllm/lib/python3.11/site-packages/pydantic/_internal/_dataclasses.py", line 121, in __init__
(APIServer pid=913608) s.__pydantic_validator__.validate_python(ArgsKwargs(args, kwargs), self_instance=s)
(APIServer pid=913608) pydantic_core._pydantic_core.ValidationError: 1 validation error for VllmConfig
(APIServer pid=913608) Assertion failed, In Mamba cache align mode, block_size (2096) must be <= max_num_batched_tokens (2048). [type=assertion_error, input_value=ArgsKwargs((), {'model_co...transfer_config': None}), input_type=ArgsKwargs]
(APIServer pid=913608) For further information visit https://errors.pydantic.dev/2.12/v/assertion_errorBefore submitting a new issue...
- Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.
extent analysis
Fix Plan
To fix the validation error, we need to ensure that max_num_batched_tokens is at least equal to block_size when Mamba cache align mode is enabled.
Here are the steps to resolve the issue:
- Increase
--max-num-batched-tokensto be at least2096when running thevllm servecommand. - Alternatively, modify the code to automatically scale
max_num_batched_tokensbased onblock_sizewhen Mamba align mode is enabled.
Example code snippet to increase --max-num-batched-tokens:
CUDA_VISIBLE_DEVICES=2,3 vllm serve Qwen3.5-122B-A10B-GPTQ-Int4 --served-model-name Qwen3.5-27B-AWQ-4bit --gpu-memory-utilization 0.9 --port 8848 -tp 2 --max-model-len 131072 --enable-auto-tool-choice --tool-call-parser qwen3_coder --reasoning-parser qwen3 --max-num-seqs 16 --enable-prefix-caching --max-num-batched-tokens 2096Verification
After applying the fix, verify that the vllm serve command runs without any validation errors related to max_num_batched_tokens and block_size.
You can check the logs for any error messages or warnings related to the validation error.
Extra Tips
- Make sure to update the
--max-num-batched-tokensvalue whenever theblock_sizechanges to ensure that the validation error does not occur again. - Consider modifying the code to automatically scale
max_num_batched_tokensbased onblock_sizewhen Mamba align mode is enabled to prevent similar issues in the future.
Vote matrix · Quick signals
Works
Did the solution work? Tap to confirm.
Easy Fix
Was it a quick fix?
Time Saver
Did it save you time?
Blocking
Was it severely blocking?
Common Issue
Are others likely hitting this too?
Flaky / Intermittent
Is it intermittent?
Verified / Reproducible
Can you reproduce it reliably?
Still need to ship something?
×6Another batch ranked right after the header list — different links, same matching logic.
TRENDING
- Feature Request: Configurable per-minute rate limiting (RPM) for models to prevent 429 errors
- Android: Hermes App + Termux install share ~/.hermes and cause silent permission loops
- hermes update emits unicode-animations ANSI demo in non-interactive logs
- hermes update downgrades aiohttp from 3.13.4 to 3.13.3
- npm install warns about deprecated @babel/plugin-proposal-private-methods
- DingTalk inbound media URLs are skipped as unreadable native image paths
- fix(dashboard): ChatPage clears header action buttons on ALL pages, not just Sessions
- [Bug]: check_web_api_key() hardcodes built-in backends — third-party web search plugins silently disabled
- Hermes Web UI 修复经验：GatewayManager 补丁、进程 D 状态、数据库升级问题
- Telegram gateway can silently drop turn after /stop with response=0 chars while internal work continues
- Bug Report: v0.14.0 上下文污染 — 历史回复碎片回注到新请求
- Bug: hermes skills search table truncates Identifier column — install fails with copied value
- [skills-index-watchdog] Skills index is stale or degraded (degraded)
- Discord approval embed not rendering on web/mobile — embed data present in API but invisible
- Idea: Discord voice-channel participation / opt-in auto-join mode
- [Feature]: Claude Code--ultrawork
- build-arm64 job deterministically fails on cold cache (Azure SAS token expires mid-build)
- [Enhancement] computer_use: action=type should fall back to key events for terminal emulators (Ghostty/Terminal.app/iTerm2)
- Feature Request: Session Recovery on Temporary Provider Outage
- [Bug]: Hermes dashboard not working on NixOS (container)
- [Feature]: Add option to ignore @all/@everyone mentions in Feishu group chats
- QQ Bot WebSocket 频繁断开：长时间工具执行阻塞 asyncio 事件循环导致心跳超时
- patch tool: new_string escape sequences (\t) get written literally
- Feature Request: i18n / 多语言支持（国际化）
- Bug: web_crawl schema lets models auto-guess "instructions" instead of asking the user via clarify
- feat: `!command` prefix for direct shell execution (like Claude Code)
- Expose currently-running cron jobs via /api/jobs (or new endpoint)
- [Bug]: Kanban parent-child handoff: scratch workspace GC destroys artifacts before child can read them
- [Bug, Windows] hermes gateway restart loses session context — planned_stop_marker not written before SIGTERM
- [Bug]: Codex→DeepSeek fallback sends assistant turns without reasoning_content → HTTP 400 (require-side cross-provider failover)
- [Bug]: Update got stuck half way, reboot it, then ModuleNotFoundError: No module named 'hermes_cli'
- Kanban dispatcher corrupt-board handling and multi-profile gateway ownership ambiguity
- Gateway can resend a short fallback message when the real final Telegram response was already delivered
- [BUG] Bedrock: Fix 'Invalid API Key format' for presigned URL tokens
- Secret redaction corrupts code syntax in tool output (write_file, execute_code, terminal)
- Unable to connect Ollama Cloud with Pro Subscription to Hermes
- feat: fuzzy substring matching for /skill autocomplete
- PRD: Autonomous market-impact prediction briefing system
- Kanban dashboard should support task/card deep links
- [Feature] Native Feishu CardKit Streaming: consolidate best-in-class implementations
- [Feature]: Inject mental model into context when using Hindsight
- Interactive CLI hides tool output despite display.tool_progress=all, and hermes chat -v does not restore it
- fix(api_server): _handle_responses drops text.format JSON schema — structured output constraints silently ignored
- state.db FTS corruption goes undetected — no integrity check, no repair path
- bug: fallback routing can select text-only models for image requests and hide the primary failure
- feat(kanban): persist worker session_id per run and pass --resume on respawn after unblock
- feat(kanban): support GitHub/OMO lifecycle bridge for Xiyou-style automation
- Expose update-safe TUI/composer hooks for voice transcript and composer events
- Hide or configure voice transcript status rows in editable dictation mode
- [Feature]: Per-Tool / Per-Toolset Approval Policies
- Context compression creates orphan sessions missing from state.db
- messaging platform
- feat: Add read-only / silent monitoring mode for WhatsApp adapter
- double-.hermes path mismatch, the HOME env var leak, and the fallback-notification UX problem
- Bug: Plattform-Bundle name `hermes-yuanbao` in `agent.disabled_toolsets` silently kills ALL tools in gateway path (Telegram + cron), CLI unaffected
- CLI /yolo (in-chat) does not bypass dangerous command approvals — env var freeze + missing enable_session_yolo call
- OpenAI Codex provider crashes with "'NoneType' object is not iterable" (HTTP None)
- DEEPSEEK_API_KEY blocked by env blocklist in gateway process — cron jobs fail with deepseek provider
- fix(feishu): Card action callback routing issues - invalid message_id and unrecognized /card command
- Discord plugin: profiles without explicit `discord:` block silently get `require_mention=true` + `auto_thread=true` (regression in cc8e5ec2a)
- [Bug]: DISCORD_ALLOWED_ROLES ignored by gateway _is_user_authorized — role-authorized users get 'Unauthorized user' rejection
- [Bug]: /new, /clear, and /reset commands freeze the terminal session
- openai-codex subscription backend returns HTTP 200 with response.output=None, causing Slack/cron failures
- RFC: Centralized Model/Provider Registry
- bug: openai-codex provider — TypeError: 'NoneType' object is not iterable on every request (gpt-5.5)
- [Feature]: Source-aware instruction gate — architectural mitigation for indirect prompt injection
- Named custom provider stale_timeout_seconds ignored because runtime provider is normalized to `custom`
- guard test (ignore)
- [Feature]: per-platform LLM request_overrides (extra_body / reasoning_effort / service_tier)
- One-shot smoke: add Flue-backed orchestration fixture
- Gateway should not treat stale Codex app-server progress as final response after post-tool silence
- `docker_run_as_host_user: true` breaks bundled skills: Hermes home is mounted into `/root/.hermes` but the container runs as a non-root user (`HOME=/home/pn`)
- [Bug]: gateway api_server streaming bypasses server-side tool-call loop when chat_template_kwargs.enable_thinking=false (model emits tool name as plain text)
- [Feature]: Pre-install python-telegram-bot in Umbrel Hermes Docker image
- YouTube Shorts filter not working in youtube-content skill
- v0.15.0 PyPI release breaks ALL platforms — plugin.yaml manifests missing from package
- RFC: On-demand tool/skill/MCP discovery — decouple schema registration from process lifecycle
- Pixshelf: local-first stock photo workflow command center
- [Bug]: baoyu infographic skill should not silently bypass image_generate
- Pixshelf v1.5: manual submission tracking for stock agencies
- `hermes config set` silently accepts unknown keys, writing them where the runtime never reads
- Honcho memory prefetch hang on fresh CLI subprocess in v0.15.0 (regression from #27190)
- [Bug] v0.15.0 Docker image: stage2-hook.sh, main-wrapper.sh missing; container_boot module removed
- Feature: Reduce cache-read token overhead for DeepSeek providers — configurable cache_ttl, skills snapshot trimming, memory compaction
- Windows: three bugs from daily use (plugin discovery, gateway exit code, Unicode decode
- holographic memory: HRR silently degrades to FTS5 when numpy is missing
- Make max_tokens configurable for aux vision calls
- Conversation compression desynchronizes session ID between agent context and gateway routing, causing silent message loss
- [Bug]: v0.15.0 Docker image:The TUI cannot be used in the dashboard.
- cron: skip_memory=True blocks fact_store/memory tools from all cron jobs
- TUI: Node.js OOM crash when agent uses browser tools repeatedly
- feat: model_profiles — per-model toolset and memory config
- Automatic background skill patching disrupts active sessions (severe impact on local models)
- ensure_hermes_home() creates root-owned dirs in profile subdirectories when kanban workers are dispatched
- Feature: opt-in webhook bypass for DISCORD_ALLOW_BOTS — allow operator-initiated probes without weakening bot-loop guard
- v0.15.0: Codex requests fail HTTP 400 when participant display_name contains non-ASCII (emoji breaks input[].name pattern)
- Architecture: State Persistence Precedence (Memory vs Skills vs Hooks)
- [Bug]: cronjob tool: create action always fails with "schedule is required for create" even when parameters are provided
- codex-oauth: 'NoneType' object is not iterable in _run_codex_stream (gpt-5.5) — every turn fails non-retryably
- Docs/Config: Plugin local scope enablement ambiguity
- [Bug]: CLI freezes after using /new command (WSL)
- Profile Codex auth can ignore global credential pool when local state is stale
- [workflow-engine] CRITICAL: variable substitution crashes on regex metachars in user input
- [workflow-engine] HIGH: loop and bash nodes leak subprocesses on timeout
- [workflow-engine] HIGH: README documents config env vars the engine never reads
- [workflow-engine] MEDIUM: workflow_run rate limit bypassable via concurrent calls (TOCTOU)
- [workflow-engine] chore: manifest gaps, side-effectful register(), dead code, unauth kanban dispatch
- [mcp_lazy] HIGH: synthetic mcp_server_<name> stub collides with a real MCP server named 'server'
- [mcp_lazy] HIGH: promote_server eager flag documented but never persisted
- [mcp_lazy] MEDIUM: _prev_mode dict leaks and goes stale; not cleared on session evict
- [mcp_lazy] MEDIUM: get_pool has unlocked check-then-set race on pool creation
- [mcp_lazy] MEDIUM: pre_tool_call gives no guidance for unpromoted server-stub calls
- [mcp_lazy] chore: undeclared pre_tool_call hook, nonexistent 'mcp_load_tools' name in docs, missing tests
- [a2a_fleet] CRITICAL: server never auto-starts — register() runs outside an event loop
- [a2a_fleet] CRITICAL: auth_required defaults to false on a cross-machine surface
- [a2a_fleet] HIGH: remove invented disable() hook — loader never calls it, port leaks on reload
- [a2a_fleet] HIGH: plugin.yaml missing kind / provides_tools / requires_env (token env undeclared)
- [a2a_fleet] MEDIUM: tighten wide-open CORS, anonymous /health peer leak, and peer-URL SSRF
- [a2a_fleet] MEDIUM: relocate tests to tests/plugins/ and cover sync-register + auth-default paths
- xai-oauth auxiliary client incorrectly uses Responses API (CodexAuxiliaryClient), causing 403 on compression/vision/web_extract
- [Bug]: Direct Copilot gpt-5.5 large resumes are killed by 12s Codex TTFB watchdog
- [Bug]: `hermes uninstall` does not work on Windows
- TUI: Thinking block leaks raw JSON and Σ character
- Hostinger VPS: migration Hermes Agent → Hermes WebUI impossible (tini + UID mismatch + sessions)
- /goal judge over-continues exploratory goals unless the assistant explicitly says the goal is complete
- /goal auto-continuation can be amplified by preflight compression/session split and resurrect stale task state
- Dashboard infinite reload loop in loopback mode — GET /api/auth/me returns 401 on every page load
- [Bug]: Provider/LLM switch leaves stale encrypted_content causing 400 errors on Telegram sessions
- [Bug]: Infinite reload loop / React state loop on Sessions tab (Firefox + Chrome) — repeated 401 on /api/auth/me (v0.15.0)
- show_reasoning should work independently of streaming in CLI mode
- Feature Request: Strip reasoning/<think> blocks from TTS preprocessing
- mcp add / mcp test raise NameError when mcp package not installed
- v0.14.0 dashboard breaks behind reverse proxies — two regressions
- Skills hub creates empty category directories when no skills installed
- [Bug]: Custom endpoint: ChatCompletions returns content, but Hermes treats response as empty (v0.14.0)
- fix: atomic_replace() fails with EXDEV when HERMES_HOME is a cross-filesystem symlink
- fix(gateway): Feishu session cancellation orphans session guard, permanently blocking messages
- Custom endpoint pricing can overestimate Crof qwen3.5-9b cost by 1,000,000x
- MCP OAuth callback: module-level port global causes port collisions and structural weaknesses vs upstream
- Bug: send_message tool bypasses validate_media_delivery_path security check
- Proposal: Add Mnemosyne to official memory provider documentation
- feat(swarm): support custom verifier/synthesizer body + skills
- Template conversion failed
- Error occurred in the operation of the agent node in the workflow.
- PubSub client overrides Sentinel client when REDIS_USE_SENTINEL is enabled
- Frontend description of the Retrieval node output does not match the actual output
- JSON type input var raise Intenal server error
- cannot extract elements from a scalar
- 负载均衡 为模型配置多组凭据，并自动调用，此功能无法选择
- add models is error
- panic: could not create filter
- Persist partially generated messages when /chat-messages/:task_id/stop is called
- MCP server connection fails with 403 — request never leaves Dify (SSRF proxy suspected)
- Support durable async execution backends for long-running workflow steps
- [Xiaomi MiMo] Credentials validation fails with 400 "Not supported model mimo-v2-flash" when using Token Plan endpoint (v0.0.7)
- After clicking preview on a parent-child segmented knowledge base, it shows 0 chunks
- Retrieval score differs between UI upload (.docx) and API upload (.txt) despite identical chunk content and embedding model
- gemini cli crash again
- Xbox gift card code damage
- Damage caused by the gemini cli crash
- ioctl(2) failed, EBADF (Bad File Descriptor)
- Feat: Support Bun as an alternative runtime/package manager for updates and extensions
- fatal error again!!!!
- ioctl error
- Critical Crash: ioctl(2) failed, EBADF in ShellExecutionService.resizePty
- ioctl(2) failed, EBADF
- v0.44.0 Regression: Critical crash with ioctl(2) failed, EBADF during PTY resize
- Crash on startup: ioctl(2) failed, EBADF in UnixTerminal.resize
- Crash: `ioctl(2) failed, EBADF` in `node-pty` during PTY resize on macOS
- Gemini CLI crashes with `ioctl(2) failed, EBADF` in `node-pty` during `resizePty`
- Remote Role
- ERROR ioctl(2) failed, EBADF /home/mich
- RangeError: Maximum call stack size exceeded
- EBADF Error during folder creationg broke session and terminal glitches
- MAIP / Gargoub Project - Mediterania - North Coast
- Gemini cli crash again in this morning
- ERROR ioctl(2) failed, EBADF
- Verified node install fails — Checksum verification failed (Cloud)
- The extended debugging key did not arrive during registration.
- CollaborationPane unmounts collaboration store on single-user instances, causing permanent "No network connection" state
- Workflow cannot be saved when the name contains "->" (Potentially malicious string)
- automation does not work and does not show an error
- Raj Ai Automation
- Default Data Loader: DOMMatrix is not defined error
- Feature: Per-node execution timestamp overlay on canvas during workflow run
- AI Agent + Vertex `gemini-3.5-flash`: 400 "missing thought_signature" on sequential multi-turn tool calls (post-#24982)
- PDF Loader in Pinecone Vector Store fails due to pdf-parse version conflict (v2 not supported)
- emailReadImap: add UID deduplication, batch size cap, and numeric uid enforcement
- Manual node execution fails with "Could not find a node" when autosave is disabled (N8N_WORKFLOWS_AUTOSAVE_DISABLED)
- Schedule Trigger stopped firing — workflow Published & active, manual executions succeed, no automated fires for 2+ hours
- [MCP SDK] create_workflow_from_code intermittently returns HTTP 500, often as a false negative (workflow persists anyway, causing duplicates on retry)
- Credential-load wedge: workflows using googleApi/jwtAuth credentials silently fail to execute after key rotation
- Google Sheets Trigger every minute is not working manual Execute is working sent email
- [BUG] Plugin marketplace MCP connector remains stuck "still connecting" when mcp-remote requires OAuth
- [redacted at user request]
- Opus 4.7 behavioral regression: loaded instruction-following discipline degraded in recent Claude Code/Cowork updates
- [BUG] Tailscale via Homebrew CLI + Mac App Store GUI, both Macs on macOS, Cowork blocked by VPN detector despite Tailscale being a mesh VPN with no traffic interception
- stopShellPty on tab switch kills active sessions (exit 143) — regression in May 27 build
- [BUG] Long URLs are broken into multiple lines and become unclickable in terminal output
- [BUG] claude rm/stop/reap SIGKILLs background session tree without SIGTERM grace, orphaning git index.lock and similar
- [BUG] Default git workflow in the system prompt was pushed without context or consent
- [MODEL] Inconsistent output quality / Ignoring instructions (overfitting and inappropriate repetition of Korean vocabulary)
- You've hit your weekly limit · resets May 31 at 5pm (Asia/Shanghai)
- Paid yearly subscription silently downgraded to Free with no user action
- [Regression v2.1.153] Plugin bash hooks fail with "echo: write error: Permission denied" on Windows (claude-mem, shell: "bash")
- [BUG] Connector toggles in conversation are not clickable — must click text label instead
- [remote-control] Input from mobile app/browser not reaching host session — output works fine
- Model fails to read/reference CLAUDE.md contents despite being loaded in context
- [BUG] Claude Desktop reinstall destroys Code chat history (transcripts + Recents) while regular Chat history, project files, and memory all survive
- Bypass mode clamps to Accept Edits even with the toggle ON (Claude Code Desktop 1.9255.2 / CC 2.1.149)
- [BUG] TUI input freezes randomly mid-typing — entire prompt becomes unresponsive for minutes
- [BUG] Cowork downloads Linux ELF binary instead of macOS binary on macOS Sonoma 14.8.7 — exit code 132 (SIGILL) on every session
- [Feature Request] Persistent project memory — sessions forget everything on close, forcing users to keep many sessions open
- [Bug] Thread context stale after sleep/resume, returns outdated date and calendar data
- [FEATURE] Add context window usage indicator and warning before auto-compaction
- [BUG] Dictation error: Invalid character in header content ["x-config-keyterms"] on Windows
- [Bug] Anthropic API Error: Server rate limiting despite normal usage
- Does delegating work to `claude -p` subprocesses reduce context accumulation in the parent session?
- [BUG] Claude Code hangs on M1 Mac when terminal says "opening browser to sign in" and browser opens
- [BUG] Claude_Preview MCP preview_start spawns dev server with main-repo cwd instead of session's worktree cwd
- [Bug] Anthropic API Error: Server rate limiting during request execution
- [Bug] Anthropic API Error: Server rate limiting on concurrent requests
- [Bug] Ultraplan ready notification fires before cloud agent completes execution
- [BUG] API 500 ERROR ALL THROUGHOUT THE DAY
- [BUG] Cowork: Live Artifacts folder path changed in 1.9255.2, no automatic migration from Documents\Claude\Artifacts
- [Bug] Auto-compact never triggers despite statusline reporting "100% context used" (v2.1.153, Max sub, 200K mode)
- [BUG] [Desktop / macOS] 'Open in → New Window' detached session: font renders smaller than main, no per-window controls, Cmd+/Cmd- keystrokes routed to main window instead
- Feature request: option to switch between classic and new minimal UI
- [Feature Request] Show timestamps for each message
- [BUG] Terminal corruption when permission prompt appears while navigating Agent Teams agent selection menu
- [FEATURE] Allow users to customize the background color of the Claude desktop app beyond the current light/dark theme presets.
- [BUG] Statusline not displaying on Windows [fixed]
- Background agent UI Stop button is a no-op for stuck agents — process keeps consuming tokens
- Background agents silently die on session pause/resume — no completion notification, no work recovery
- Add option to hide email address from welcome banner
- [BUG] SSH Remote: `projects` field in remote ~/.claude.json becomes null after desktop restart — jsonl files intact, UI shows 'No messages yet' for every session
- [Bug] Claude Code not applying fixes despite claiming to complete tasks
- billing is unfair and poorly documented
- [BUG] Claude Code on the web: declared plugins inactive on first session, require restart to fully load
- [BUG] Restore from archive deleted sessions instead of restoring them
- [BUG] M365 connector fails with AADSTS50011 in Cowork — localhost vs 127.0.0.1 redirect URI mismatch
- claude agents: workflow slash-commands missing from dispatch-input completion (regression-adjacent to #61424)
- Claude Desktop's Info.plist missing TCC usage strings, blocks all EventKit-based MCP servers
- False-positive safety blocks on self-administered governance amendments — request for owner-authority mode for verified professional users
- [BUG] Stop pushing "AUTO"-mode
- [DOCS] Plugin marketplace guide omits `skipLfs` option for git-based sources
- [DOCS] MCP docs omit combined startup notification for MCP server and connector authentication
- [DOCS] Agent view docs omit macOS Privacy & Security identity for background agents
- [DOCS] Npm update docs do not explain release-channel behavior for `claude update`
- [DOCS] Agent SDK docs omit `subagent_type: "claude"` worktree and output persistence behavior
- [DOCS] Background session docs omit `$CLAUDE_JOB_DIR` temp-file behavior
- [FR] mask env-var values in 'claude mcp get <server>' output
- [FR] subagent worktrees should not inherit stale local 'user.email' from prior dispatches
- [BUG] Windows: Grep tool leaks rg.exe + conhost.exe processes (~2000 zombies / 14 GB RAM in long sessions)
- [BUG] Stats dashboard "Peak hour" appears off by one hour
- [BUG] Diff highlight (teal SGR background) bleeds past changed text in 2.1.150–2.1.153
- [FEATURE] confirm before deleting session
- Plugin PostToolUse hooks still silently skip in Claude Desktop / Cowork (re-filing closed #51904)
- /code-review skill: silent fallback to main...HEAD reviews other people's commits, and JSON-only output is hard to read
- Monitor tool doesn't source the shell snapshot like Bash does; PATH-dependent tools (jq, sleep, etc.) fail in Monitor commands on macOS/Nix
- [Bug] Long input lines truncated with ellipsis while typing instead of wrapping in terminal UI
- [FEATURE] VS Code extension: Render submitted user messages as Markdown in chat
- OSC 52 copy from Claude TUI doesn't reach clipboard inside tmux (regression in 2.1.146–2.1.153)
- [BUG] RemoteTrigger create/update returns HTTP 400 with circular error: "event_type is required" / "unknown field event_type"
- [BUG] Option to hide or minimize the built-in "status footer" (multi-line debug/cost panel) [re-raise of #31475]
- [Bug] Feedback submissions being closed without review or action
- [FEATURE] Word-jump cursor navigation in Chat input (option+arrow / bindable actions)
- [FEATURE] ! shell mode: filesystem tab completion
- [BUG] API Error: Usage credits required for 1M context
- claude agents: OSC 52 clipboard emission broken in tmux (regression in 2.1.146–2.1.153)
- CLI crashes on macOS 15 M3 - exit code 1
- [FEATURE] Support Cmd+V image paste from clipboard
- [FEATURE] Enhance claude.ai M365 connector to support MS Planner
- [BUG] Slash command autocomplete hijacks pasted absolute file paths starting with /
- PreToolUse hook `if` filter false-positives on complex Bash commands
- [BUG] Diff panel hangs/whites out
- Feature Request: Support drag-and-drop for binary documents (.wps, .doc, .docx, .xlsx, .pdf) in VS Code extension
- [BUG] activation of 1M context in VSCode
- [FEATURE] Support i18n / language localization for built-in slash command outputs
- Ctrl+V para colar imagens deixou de funcionar no CLI (Windows, PowerShell)
- [FEATURE] Please add Norwegian (Bokmål/Nynorsk) language support to the Claude Code interface
- [BUG] OTel log events (claude_code.user_prompt, api_request_body, tool_decision, hook_execution_complete) emitted with empty trace_id/span_id while sibling spans correlate correctly
- [BUG] Cowork crashes on every message, no VM logs generated, missing AppData\Roaming\Claude
- [FEATURE] first-class session handoff + per-session token budgets for unattended runs
- [FEATURE] Smart paste: convert clipboard code to file reference chips (like Cursor)
- [Feature Request] Restore chat pin functionality to title chat submenu
- [BUG] SIGILL issues with version 2.1.153
- [BUG] Cowork plugin upload fails with generic "Plugin validation failed" when a `description` field in any SKILL.md frontmatter contains angle brackets (`<…>`)
- [BUG] Desktop App 2.1.144+: startup scanner deletes cliSessionId from claude-code-sessions local files on every launch — session not found on disk
- [Feature Request] Add keyboard shortcut to copy last message with proper formatting
- [MODEL] Opus 4.7 not 1M
- Allow naming/renaming background agents in `claude agents` view
- Stale worktrees in .claude/worktrees/ are never cleaned up, consuming massive disk space
- Agent worktrees are never cleaned up, silently consuming disk space
- Subagent worktrees not auto-cleaned when reviewer writes scratch files
- [Bug] Skill initialization hangs for extended duration in Plan Mode
- Claude Desktop writes malformed registry Run entry (nested escaped quotes) - crashes Windows Task Manager and other Run-key parsers
- IME candidate window shows at bottom-right corner instead of caret position (Windows CMD)
- [BUG] Pressing 'Escape' doesn't close the /BTW conversation when the main conversation is asking for approval
- [BUG] Opus 4.7 (1M) intermittently emits empty-string values for tool_use.input fields, killing the session
- FleetView agent UI shows "running" with incrementing elapsed time after agent has returned
- /doctor flags context-scoped cmd+c binding as macOS conflict (false positive)
- [BUG] Text Rendering in Elvish
- Desktop app: Bypass Permissions mode flips to Accept Edits on first prompt (M5 / macOS 26.5)
- [Workaround] Date-Weekday Verification Hook — Prevents Claude from writing wrong weekdays
- [BUG] Claude Code create c:/memfs directory without asking me.
- [BUG] Claude Code's Bash execution waits forever with no processes running
- [BUG] usage stays stuck waiting for 5 hr limit after upgrading to premium seat in team plan
- [Workflow tool] resume cache is unreachable for nontrivial workflows because LLM dispatchers can't transcribe args byte-exactly
- Code review (Preview): "Add a repository" shows no results for private GitHub org repos
- [BUG] /context commands blows up context
- [Feature Request] Add precache expiry hook to enable proactive compaction before token eviction
- [BUG] Context indicator shows 0% at session start despite ~20K+ tokens already loaded
- [Feature Request] Add semantic search for --resume session history
- [Feature Request] Add session search, tagging, and filtering capabilities
- [BUG] Cowork Dispatch reports "desktop not available" on Windows 11 while standard Cowork works normally
- [Bug] Claude Code provides incorrect suggestions with high confidence despite errors
- defaultMode: acceptEdits silently overrides per-path permissions.ask rules for Write/Edit
- [FEATUR configurable tip interval (e.g. tipIntervalSeconds: 30 in settings)E]
- Plugin marketplace fails to load: schema rejects 'displayName' key (v2.1.153)
- claude agents: in-session copy uses broken OSC 52 path while overview correctly uses tmux buffer
- [BUG] Plugin agent descriptions (and custom agents) load unconditionally into context — no parity with disable-model-invocation for skills
- Crashed ultrareview consumed a free credit despite producing zero findings
- [Bug] Character rendering issue - invisible or missing text display
- [BUG] Cowork: processo Claude Code encerra com código 3 — .claude.json não contém token de autenticação (Windows 11 25H2)
- [BUG] 2.1.153 silently discards tools/list response from rmcp 0.12.0 HTTP MCP server (works in 2.1.152, wire-identical handshake)
- VS Code extension: option to auto-resume last session when reopening a workspace folder
- [Bug] Conversation continuation failure
- [BUG] Cowork crashes every time I start a new chat or attempt to continue an existing one in any project. The error displayed is: "Claude Code è andato in crash
- [Bug] Unannounced quota changes
- Native update/install fails with 'socket connection was closed unexpectedly' behind proxy — undici TLS incompatibility
- [BUG] Session name reverting after manual change
- [BUG] 非正常思考，上下文过长时，一直显示思考，点击interrupt按钮失效
- Honor `tools:` frontmatter when an agent is invoked via `@mention` — strip `Task` only when the agent did not declare it
- macOS TCC popup still recurring on v2.1.153 — "2.1.153" would like to access data from other apps
- Claude Code leaks pty handles — exhausts pseudo-terminals on macOS after long session
- [Bug] Agent fails to execute or respond to user input
- [BUG] Persistent "Expecting value: line 1 column 1 (char 0)" JSON parse error after tool execution
- [Feature Request] Implement proactive unit test coverage recommendations for recurring bugs
- VS Code panel lacks status line + terminal lacks image paste in Codespaces, forcing a tradeoff
- `/powerup` only shows ~10 lessons — allow viewing the full catalog
- [Bug] Context contamination after auto-compact with unrelated email draft of Tejo/Sado Basin
- [Bug] VSCode terminal output displays corrupted text with garbled symbols
- [Feature Request] Add LaTeX/KaTeX math rendering to TUI
- [Bug] Sub-agent PR review results not validated by orchestrating agent
- Subagents on Pro 1M tier: trivial probes pass, real workloads fail at first tool call (probe-vs-workload divergence)
- Path-scoped rules and subdirectory CLAUDE.md not loaded when creating new files matching the pattern
- AskUserQuestion: cancelling during extended thinking poisons the whole session with 400 'thinking blocks cannot be modified' (2.1.153); concurrent prompts overwrite each other
- Ideas Missing from Claude Cowork Menu (Windows)
- [BUG_BOUNTY_SAFE_POC_2026] Prompt Injection RCE Test - Command Execution Proof
- [BUG] Cowork scheduled task: execution history row not showing after successful run
- Resuming an extended-thinking session fails permanently with 400 "thinking blocks cannot be modified" (transcript stores thinking text as empty but keeps signature)
- [Bug] Plugin-registered CwdChanged and FileChanged hooks don't fire (settings.json works) — v2.1.153
- Auto-archive on PR merge / branch delete — clarify autoArchiveSessions semantics or add dedicated opt-out
- `claude mcp add` echoes Authorization header value verbatim to stdout, leaks bearer tokens to terminal and session transcripts
- [BUG] Bug report — /insights skill, Claude Code The /insights skill outputs a malformed file path.
- Plugin slash commands render with '*'-inline format instead of two-column, despite matching official plugin shape
- [Bug] Unexpected long text generation without user input or goal
- [Bug] Thinking blocks causing task progression blocked without user modification
- [BUG] (Critical!) contamination by an unknown session simirlar to the report => [Bug] Context contamination after auto-compact with unrelated email draft of Tejo/Sado Basin #63137
- [Critical] Opus 4.7 Korean output degeneration — Korean grammar itself collapses in long contexts
- [BUG] Title: Autocompact buffer persists across /clear — wastes tokens for irrelevant old context
- [Bug] Auto-Compact loses user input before processing in conversation history
- Feature: per-invocation effort parameter + runtime session-config introspection for skills
- Auto-mode classifier mislabels Azure DevOps vote -5 as "Reject" when denying PR vote actions
- [BUG] Claude Desktop and Claude Code CLI never re-register MCP tools after OAuth 2.1 handshake on a remote HTTP server
- [BUG] Workspace file tags leak across sessions
- [BUG] Ink renderer crashes on Windows 11 build 26200 (Canary) duplicate banners, terminal mode leaks, mid-operation aborts
- [BUG] Claude Code Desktop issue
- PTY master fd leak in Claude desktop app exhausts macOS kern.tty.ptmx_max after ~2-3 days
- [BUG] Claude Code — Session Management after Unexpected Interruption
- [Windows] Cowork OpenTelemetry exporter does not initialize - zero events emitted to any destination, including loopback
- [Bug] Opus 4.7: 400 `thinking blocks ... cannot be modified` on long extended-thinking sessions, triggered by history-altering events (scheduled prompts / parallel tool-call cancellation)
- [BUG] API Error: Server is temporarily limiting requests (not your usage limit) · Rate limited
- Multi-plugin custom marketplace: only first plugin registered in installed_plugins.json, skills don't load
- [BUG] Git push through the SDK's git proxy fan-outs into ~500 GitHub REST API calls, exhausting the 5,000/hour budget after a handful of pushes
- [BUG] Claude took liberties it really shouldn't with my global config
- [BUG] Agent window focus lost after navigating with arrow keys, causing scroll deadlock
- [BUG] `--model` flag silently ignored in interactive sessions (works in `--print` only)
- [BUG] Dispatch permanently shows "desktop appears offline" on Windows 11 - never worked on first use
- feat: support per-command enableWeakerNetworkIsolation as safer alternative to dangerouslyDisableSandbox
- /code-review outputs a raw JSON array instead of readable findings
- [BUG] Cowork — Additional allowed domains ignored on Team plan; same domain works on Pro plan
- Haiku
- [Bug] False positive blocking beneficial outcomes in tool execution
- 3P Bedrock SSO: credentials silently expire without triggering re-auth on day 2+
- CLAUDE_AUTOCOMPACT_PCT_OVERRIDE in settings.json env block silently ignored by autocompact logic
- Auto-compaction deletes main session JSONL before verifying summary completion, causing data loss
- [Bug] Claude Code not executing stated actions or producing expected results
- [FEATURE] Deferred Messages — Queue Input for End of Turn
- [BUG] Up/Down arrows in input box navigate history instead of moving cursor — regression in 2.1.149+
- Cancelling a parallel tool-call batch corrupts thinking blocks -> 400 "thinking blocks cannot be modified" permanently wedges the session
- Claude Code caused data loss, then contradicted itself about recovery (two incidents, one session)
- [Bug] Unclear error messages from Claude Code CLI
- [Bug] Agent tool rejecting due to context size limit exceeded
- claude agents: daemon and bg-spare processes spin at ~100% CPU when idle
- [BUG] Compaction fails with "context window limit" error even when context usage is low (e.g., 20%) — regression in v2.1.153
- Remote Control entitlement lost after May 27-28 incident — `Error: Remote Control is not yet enabled for your account` on active Max subscription
- PreToolUse hook exit code 2 does not block Write tool
- [Bug] Thinking blocks in latest assistant message are immutable
- GUI: dispatch file:// and custom-scheme clicks to OS shell handler
- Show current model in statusLine by default
- [Bug] Agent console becomes unresponsive to keyboard input after multiple agents initialized
- [FEATURE] PreToolUse hooks should have a way of updating the environment
- [Bug] Unable to start or use Claude Code CLI
- [BUG] Repository not visible in Claude Code web repo picker
- Session permanently wedged on 400 "thinking blocks cannot be modified" after parallel tool_results
- [Bug] @ autocomplete loses sibling repos after a file edit in multi-repo workspace
- Unclear error message when creating sub-agent without authentication
- [Bug] Anthropic API errors causing frequent failures and high token usage
- [BUG] @ mention file picker only shows packages, not individual files (desktop app - Code tab)
- [Bug] TUI panel footer remains sticky and consumes excessive terminal space
- PR-status polling exhausts GitHub GraphQL rate limit on repos with many open PRs
- [BUG] Windows: welcome panel not shown in some project folders (2.1.153)
- [Bug] Anthropic API Error: thinking blocks corrupted during context compaction with extended thinking enabled
- API 400 "thinking blocks cannot be modified" permanently bricks session during agent activation (interleaved thinking + tool use)
- Right-click Copy copies the whole message instead of the selection; pasted text retains dark background
- Mid-session model switch corrupts conversation when extended thinking is enabled (API 400: 'thinking blocks cannot be modified')
- [BUG] Markdown file links in chat output do not open files when clicked (VS Code extension)
- Stuck retry loop: `400 thinking blocks cannot be modified` on large interleaved-thinking turns using AskUserQuestion
- [FEATURE] Prompt user for approval before auto-compaction proceeds
- Custom MCP connectors not attachable to scheduled routines — no UUID discovery path
- [BUG] Claude in Chrome — Navigation blocked for teams.cloud.microsoft and outlook.cloud.microsoft after Microsoft domain migration**
- [BUG] Claude Desktop — Personal plugins panel renders list but is entirely non-interactive (macOS, v1.9255.2)
- [Bug] error when using Workflows
- [BUG] Persistent "update available" notification despite being on latest version
- [BUG] Sweep Agent from /code-review never completes
- [Bug] Tool calls not executing or returning results
- [FEATURE] Cloud-synced memory and settings across machines
- [Bug] Terminal UI freezes when Ctrl+O view exits during interactive prompt in plan mode
- Continuous api errors when using claude code with Opus 4.7 with thinking on low
- [Feature Request] Add support for installing and using previous Claude Code versions
- [Bug] Extended Thinking: Summarized thinking blocks fail signature validation when resent to API
- [Bug] Anthropic API Error: 'thinking' blocks cannot be modified
- [Bug] Anthropic API Error: Thinking blocks cannot be modified with extended thinking mode
- Feature request: Lazy/on-demand MCP server connections
- [Bug] Tool Arguments Parsed as String Instead of Object
- [Bug] Anthropic API Error: Insufficient context provided
- [Bug] Claude Opus occasionally uses moskovian(russian) orthography instead of Ukrainian in system-prompted responses
- Opus 4.8: backgrounded task completions (subagents AND Bash) crash with 400 "thinking blocks cannot be modified"
- [Bug] Opus 4.7 fabricates stable preferences ("my default") to rationalize arbitrary choices when challenged
- [Bug] Unable to update Claude Code CLI
- [BUG] Desktop app: /remote-control mints link + connects bridge (main.log) but in-chat link/QR panel never renders
- Feature: sessionColor and sessionName in .claude/settings.json
- [BUG] Anthropic API error: thinking blocks
- [FEATURE] Support Remote MCPs in Cowork as in Claude Code
- [Bug] Anthropic API Error: 400 Bad Request with Redacted Thinking - 0 4.7 & 4.8
- [Bug] Anthropic API Error: Cannot modify thinking blocks from different model versions
- Interleaved thinking + multi-tool turn corrupts thinking block (text blanked, signature kept) → permanent 400 'blocks must remain as they were'
- [BUG] Mode/permission changes mid-tool-loop (effortLevel: xhigh) poisons entire session
- Session failure log: Opus 4.6 ignores its own rules for an entire session
- [BUG] "400 Guardrail was enabled" error when using Claude Opus 4.8 with AWS Bedrock
- [Feature Request] Add subagent approach selection option to avoid accidental feedback
- Persistent 400 'thinking blocks in the latest assistant message cannot be modified' — interleaved thinking persisted with empty text + signature bricks sessions
- [BUG] DesktopvsApp
- [BUG] Opus 4.7 cache hit rate collapse after May 27 incident — Messages 1.1k→88.9k in 9 minutes, $630/session
- [Bug] Anthropic API Error: Invalid thinking block format
- [BUG] FUCK CLAUDE
- Opus 4.8 extended thinking: Stop hook block re-entry corrupts thinking blocks → 400
- [Bug] 4.8 Fails when accessing previous model history
- [Bug] Unintended File Modifications During Execution
- [DOCS] Model configuration docs omit lean system prompt default scope and model exceptions
- Add "Always allow globally" option to permission prompts
- Server-side model upgrade (Opus 4.7→4.8) wedges in-flight sessions with `thinking blocks cannot be modified` 400
- [DOCS] AskUserQuestion docs missing multiple-choice prompt decision threshold
- [DOCS] Agent view docs omit shell-command background session launch syntax
- [DOCS] Agent view dispatch input docs incorrectly imply `/logout` dispatches as a prompt
- [DOCS] Claude in Chrome docs omit connected-browser selection behavior
- [DOCS] Plugin docs omit `defaultEnabled: false` for opt-in plugins
- Feature Request: Customizable chat text colors for user and assistant messages
- [DOCS] `/plugin` Discover tab docs omit directory-based suggested plugin pins
- VSCode Chrome integration silently fails: 3 distinct bugs
- [DOCS] MCP stdio docs omit session environment variables
- [Bug] Anthropic API error on second request within session with Claude Opus 4.8
- Cowork emits a blank session "index" handoff on focus when a CLI session is paused awaiting input
- [DOCS] MCP docs omit `claude mcp list/get` pending-approval output for unapproved project servers
- [BUG] /compact fails with 400 error when last assistant turn contains thinking blocks
- [DOCS] `/claude-api` docs omit Opus 4.8 migration guidance
- [DOCS] Fast mode docs still recommend deprecated Opus 4.6 override variable
- [DOCS] Bash tool docs omit `$TMPDIR` consistency across sandboxed and unsandboxed commands
- [Bug] Anthropic API Error: 400 Bad Request on Extended Thinking
- [DOCS] Background session docs omit worktree-isolation behavior for spawned subagents
- Built-in mechanistic self-verification of verifiable claims (symmetric to the auto permission gate)
- [DOCS] Worktree docs do not clarify `worktree.baseRef: "head"` inside linked worktrees
- [BUG] Excessive RAM usage with multiple parallel chats (~10 sessions → 30 GB memory pressure, macOS OOM)
- [DOCS] Managed MCP policy docs omit invalid `allowedMcpServers`/`deniedMcpServers` entry behavior
- [DOCS] Effort docs omit `CLAUDE_CODE_ALWAYS_ENABLE_EFFORT` unsupported-model behavior
- Regression (2.1.147–2.1.150?): resuming an extended-thinking session after a CC update/model-switch → unrecoverable 400, session bricked
- [DOCS] Windows updater docs omit `claude.exe` in-use recovery guidance
- [DOCS] VS Code auto mode docs still tie mode-picker visibility to bypass-permissions setting
- [DOCS] MCP docs omit `/mcp` tool list and detail rendering behavior
- [DOCS] Fine-grained tool streaming docs still describe provider opt-in behavior
- bypassPermissions: session startup reads flat pref, GUI toggle writes per-account pref — they never sync
- [BUG] Claude Desktop Code tab causes disk write limit violation — 8.5GB in 11 min, macOS kills app (M5, v1.9659.1)
- Ultrareview v2.1.96: docs describe /tasks command + claude ultrareview --json subcommand that don't exist; findings hard to read after completion
- I'd be happy to help create a GitHub issue title, but I don't see the error message in your message. Could you please share the specific error you're encountering? That way I can generate an accurate and descriptive issue title for you.
- [BUG] Claude in Chrome `file_upload` rejects all scheduled-task sessions with misleading error (real cause: INVALID_SESSION)
- Extended thinking: signed thinking block 'cannot be modified' (400) permanently wedges session
- RTL text support for Hebrew (and Arabic) in Claude Code
- [Bug] Random errors occurring across multiple operations