Issue Station
Sign In
Tools hereToolbox

vllm - ✅(Solved) Fix [Bug]: Gibberish output and collapsing generation throughput with Qwen3.5-35B-A3B-FP8 and speculative decoding enabled [1 pull requests, 1 participants]

Official PRs (…)
ON THIS PAGE

Recommended Tools

×6

Utilities 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

Loading…
GitHub stats
vllm-project/vllm#36872•Fetched 2026-04-08 00:43:51
View on GitHub
Comments
0
Participants
1
Timeline
7
Reactions
2
Author
marionchadalmarionchadal
Participants
marionchadalmarionchadal
Timeline (top)
subscribed ×3cross-referenced ×2labeled ×1referenced ×1

When serving Qwen/Qwen3.5-35B-A3B-FP8 with MTP speculative decoding enabled, the model progressively produces incoherent tokens (random Unicode characters, repeated symbols, gibberish). The model is used in a RAG engine with text and images as inputs.

Error Message

During multi-turn chat completions with tool calling, the output quality degrades across consecutive requests. The speculative decoding metrics reflect this degradation: Request 1 (output is coherent):

SpecDecoding metrics: Mean acceptance length: 2.23, Avg Draft acceptance rate: 61.3%
Per-position acceptance rate: 0.798, 0.427

Root Cause

When serving Qwen/Qwen3.5-35B-A3B-FP8 with MTP speculative decoding enabled, the model progressively produces incoherent tokens (random Unicode characters, repeated symbols, gibberish). The model is used in a RAG engine with text and images as inputs.

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): 15 On-line CPU(s) list: 0-14 Vendor ID: AuthenticAMD Model name: AMD EPYC 9124 16-Core Processor CPU family: 25 Model: 17 Thread(s) per core: 1 Core(s) per socket: 1 Socket(s): 15 Stepping: 1 BogoMIPS: 5999.99 Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm rep_good nopl cpuid extd_apicid tsc_known_freq pni pclmulqdq ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm cmp_legacy svm cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw perfctr_core ssbd ibrs ibpb stibp vmmcall fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves avx512_bf16 clzero xsaveerptr wbnoinvd arat npt lbrv nrip_save tsc_scale vmcb_clean pausefilter pfthreshold v_vmsave_vmload vgif avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid overflow_recov succor fsrm arch_capabilities Virtualization: AMD-V Hypervisor vendor: KVM Virtualization type: full L1d cache: 960 KiB (15 instances) L1i cache: 960 KiB (15 instances) L2 cache: 7.5 MiB (15 instances) L3 cache: 240 MiB (15 instances) NUMA node(s): 1 NUMA node0 CPU(s): 0-14 Vulnerability Gather data sampling: Not affected Vulnerability Indirect target selection: Not affected Vulnerability Itlb multihit: Not affected Vulnerability L1tf: Not affected Vulnerability Mds: Not affected Vulnerability Meltdown: Not affected Vulnerability Mmio stale data: Not affected Vulnerability Reg file data sampling: Not affected Vulnerability Retbleed: Not affected Vulnerability Spec rstack overflow: Vulnerable: Safe RET, no microcode Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization Vulnerability Spectre v2: Mitigation; Retpolines; IBPB conditional; IBRS_FW; STIBP disabled; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected Vulnerability Srbds: Not affected Vulnerability Tsa: Vulnerable: Clear CPU buffers attempted, no microcode Vulnerability Tsx async abort: Not affected Vulnerability Vmscape: Not affected

PR fix notes

PR #36910: [WIP] [BugFix] Forward spec_step_idx in MTP wrappers and eagle proposer/speculator

Description (problem / solution / changelog)

Purpose

Fix spec_step_idx not being forwarded from MTP wrapper classes (Qwen3NextMTP, Qwen3_5MTP) to the inner MultiTokenPredictor, and not being passed by the eagle proposer/speculator loops.

MTP models with num_mtp_layers > 1 use spec_step_idx % num_mtp_layers to select which decoder layer to use for each speculative step. Two issues caused the wrong layer to always be selected:

  1. Wrapper classes swallow spec_step_idx: Qwen3NextMTP.forward() and Qwen3_5MTP.forward() accept **kwargs but never forward spec_step_idx to the inner model's forward(), which expects it.

  2. Eagle proposer/speculator never pass spec_step_idx: The proposer loop and speculator loop never include spec_step_idx in model_kwargs or compute_logits calls, so every draft token beyond the first silently uses layer 0 instead of the correct layer.

This follows the existing correct pattern used by DeepSeekMTPModel, ExaoneMoeMTP, and other MTP implementations that explicitly accept and forward spec_step_idx.

Fixes #36872

Test Plan

Test Result

Changed files

  • vllm/model_executor/models/qwen3_5_mtp.py (modified, +7/-1)
  • vllm/model_executor/models/qwen3_next_mtp.py (modified, +7/-1)
  • vllm/v1/spec_decode/eagle.py (modified, +9/-3)
  • vllm/v1/worker/gpu/spec_decode/eagle/speculator.py (modified, +5/-2)

Code Example

==============================
        System Info
==============================
OS                           : Ubuntu 24.04.4 LTS (x86_64)
GCC version                  : (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0
Clang version                : Could not collect
CMake version                : Could not collect
Libc version                 : glibc-2.39

==============================
       PyTorch Info
==============================
PyTorch version              : 2.10.0+cu126
Is debug build               : False
CUDA used to build PyTorch   : 12.6
ROCM used to build PyTorch   : N/A

==============================
      Python Environment
==============================
Python version               : 3.12.3 (main, Mar  3 2026, 12:15:18) [GCC 13.3.0] (64-bit runtime)
Python platform              : Linux-6.8.0-101-generic-x86_64-with-glibc2.39

==============================
       CUDA / GPU Info
==============================
Is CUDA available            : True
CUDA runtime version         : Could not collect
CUDA_MODULE_LOADING set to   : 
GPU models and configuration : GPU 0: NVIDIA L40S
Nvidia driver version        : 580.126.20
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):                                  15
On-line CPU(s) list:                     0-14
Vendor ID:                               AuthenticAMD
Model name:                              AMD EPYC 9124 16-Core Processor
CPU family:                              25
Model:                                   17
Thread(s) per core:                      1
Core(s) per socket:                      1
Socket(s):                               15
Stepping:                                1
BogoMIPS:                                5999.99
Flags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm rep_good nopl cpuid extd_apicid tsc_known_freq pni pclmulqdq ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm cmp_legacy svm cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw perfctr_core ssbd ibrs ibpb stibp vmmcall fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves avx512_bf16 clzero xsaveerptr wbnoinvd arat npt lbrv nrip_save tsc_scale vmcb_clean pausefilter pfthreshold v_vmsave_vmload vgif avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid overflow_recov succor fsrm arch_capabilities
Virtualization:                          AMD-V
Hypervisor vendor:                       KVM
Virtualization type:                     full
L1d cache:                               960 KiB (15 instances)
L1i cache:                               960 KiB (15 instances)
L2 cache:                                7.5 MiB (15 instances)
L3 cache:                                240 MiB (15 instances)
NUMA node(s):                            1
NUMA node0 CPU(s):                       0-14
Vulnerability Gather data sampling:      Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Not affected
Vulnerability Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Not affected
Vulnerability Spec rstack overflow:      Vulnerable: Safe RET, no microcode
Vulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:                Mitigation; Retpolines; IBPB conditional; IBRS_FW; STIBP disabled; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Vulnerable: Clear CPU buffers attempted, no microcode
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Not affected

==============================
Versions of relevant libraries
==============================
[pip3] flashinfer-python==0.6.4
[pip3] numpy==2.2.6
[pip3] nvidia-cublas-cu12==12.6.4.1
[pip3] nvidia-cuda-cupti-cu12==12.6.80
[pip3] nvidia-cuda-nvrtc-cu12==12.6.77
[pip3] nvidia-cuda-runtime-cu12==12.6.77
[pip3] nvidia-cudnn-cu12==9.10.2.21
[pip3] nvidia-cudnn-frontend==1.18.0
[pip3] nvidia-cufft-cu12==11.3.0.4
[pip3] nvidia-cufile-cu12==1.11.1.6
[pip3] nvidia-curand-cu12==10.3.7.77
[pip3] nvidia-cusolver-cu12==11.7.1.2
[pip3] nvidia-cusparse-cu12==12.5.4.2
[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==12.6.85
[pip3] nvidia-nvshmem-cu12==3.4.5
[pip3] nvidia-nvtx-cu12==12.6.77
[pip3] pyzmq==27.1.0
[pip3] torch==2.10.0+cu126
[pip3] torch_c_dlpack_ext==0.1.5
[pip3] torchaudio==2.10.0+cu129
[pip3] torchvision==0.25.0+cu129
[pip3] transformers==4.57.6
[pip3] triton==3.6.0
[conda] Could not collect

==============================
         vLLM Info
==============================
ROCM Version                 : Could not collect
vLLM Version                 : 0.17.1
vLLM Build Flags:
  CUDA Archs: Not Set; ROCm: Disabled
GPU Topology:
  	GPU0	CPU Affinity	NUMA Affinity	GPU NUMA ID
GPU0	 X 	0-14	0		N/A

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

==============================
     Environment Variables
==============================
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
TORCHINDUCTOR_CACHE_DIR=/tmp/torchinductor_ubuntu

---

vllm serve Qwen/Qwen3.5-35B-A3B-FP8 \
  --port x \
  --gpu-memory-utilization 0.95 \
  --max-model-len 62144 \
  --tensor-parallel-size 1 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}'

---

SpecDecoding metrics: Mean acceptance length: 2.23, Avg Draft acceptance rate: 61.3%
Per-position acceptance rate: 0.798, 0.427

---

SpecDecoding metrics: Mean acceptance length: 1.02, Avg Draft acceptance rate: 0.9%
Per-position acceptance rate: 0.018, 0.000

---

SpecDecoding metrics: Mean acceptance length: 1.00, Avg Draft acceptance rate: 0.0%
Per-position acceptance rate: 0.000, 0.000
RAW_BUFFERClick to expand / collapse

Your current environment

<details> <summary>The output of <code>python collect_env.py</code></summary> 
==============================
        System Info
==============================
OS                           : Ubuntu 24.04.4 LTS (x86_64)
GCC version                  : (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0
Clang version                : Could not collect
CMake version                : Could not collect
Libc version                 : glibc-2.39

==============================
       PyTorch Info
==============================
PyTorch version              : 2.10.0+cu126
Is debug build               : False
CUDA used to build PyTorch   : 12.6
ROCM used to build PyTorch   : N/A

==============================
      Python Environment
==============================
Python version               : 3.12.3 (main, Mar  3 2026, 12:15:18) [GCC 13.3.0] (64-bit runtime)
Python platform              : Linux-6.8.0-101-generic-x86_64-with-glibc2.39

==============================
       CUDA / GPU Info
==============================
Is CUDA available            : True
CUDA runtime version         : Could not collect
CUDA_MODULE_LOADING set to   : 
GPU models and configuration : GPU 0: NVIDIA L40S
Nvidia driver version        : 580.126.20
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):                                  15
On-line CPU(s) list:                     0-14
Vendor ID:                               AuthenticAMD
Model name:                              AMD EPYC 9124 16-Core Processor
CPU family:                              25
Model:                                   17
Thread(s) per core:                      1
Core(s) per socket:                      1
Socket(s):                               15
Stepping:                                1
BogoMIPS:                                5999.99
Flags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm rep_good nopl cpuid extd_apicid tsc_known_freq pni pclmulqdq ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm cmp_legacy svm cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw perfctr_core ssbd ibrs ibpb stibp vmmcall fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves avx512_bf16 clzero xsaveerptr wbnoinvd arat npt lbrv nrip_save tsc_scale vmcb_clean pausefilter pfthreshold v_vmsave_vmload vgif avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid overflow_recov succor fsrm arch_capabilities
Virtualization:                          AMD-V
Hypervisor vendor:                       KVM
Virtualization type:                     full
L1d cache:                               960 KiB (15 instances)
L1i cache:                               960 KiB (15 instances)
L2 cache:                                7.5 MiB (15 instances)
L3 cache:                                240 MiB (15 instances)
NUMA node(s):                            1
NUMA node0 CPU(s):                       0-14
Vulnerability Gather data sampling:      Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Not affected
Vulnerability Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Not affected
Vulnerability Spec rstack overflow:      Vulnerable: Safe RET, no microcode
Vulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:                Mitigation; Retpolines; IBPB conditional; IBRS_FW; STIBP disabled; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Vulnerable: Clear CPU buffers attempted, no microcode
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Not affected

==============================
Versions of relevant libraries
==============================
[pip3] flashinfer-python==0.6.4
[pip3] numpy==2.2.6
[pip3] nvidia-cublas-cu12==12.6.4.1
[pip3] nvidia-cuda-cupti-cu12==12.6.80
[pip3] nvidia-cuda-nvrtc-cu12==12.6.77
[pip3] nvidia-cuda-runtime-cu12==12.6.77
[pip3] nvidia-cudnn-cu12==9.10.2.21
[pip3] nvidia-cudnn-frontend==1.18.0
[pip3] nvidia-cufft-cu12==11.3.0.4
[pip3] nvidia-cufile-cu12==1.11.1.6
[pip3] nvidia-curand-cu12==10.3.7.77
[pip3] nvidia-cusolver-cu12==11.7.1.2
[pip3] nvidia-cusparse-cu12==12.5.4.2
[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==12.6.85
[pip3] nvidia-nvshmem-cu12==3.4.5
[pip3] nvidia-nvtx-cu12==12.6.77
[pip3] pyzmq==27.1.0
[pip3] torch==2.10.0+cu126
[pip3] torch_c_dlpack_ext==0.1.5
[pip3] torchaudio==2.10.0+cu129
[pip3] torchvision==0.25.0+cu129
[pip3] transformers==4.57.6
[pip3] triton==3.6.0
[conda] Could not collect

==============================
         vLLM Info
==============================
ROCM Version                 : Could not collect
vLLM Version                 : 0.17.1
vLLM Build Flags:
  CUDA Archs: Not Set; ROCm: Disabled
GPU Topology:
  	GPU0	CPU Affinity	NUMA Affinity	GPU NUMA ID
GPU0	 X 	0-14	0		N/A

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

==============================
     Environment Variables
==============================
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
TORCHINDUCTOR_CACHE_DIR=/tmp/torchinductor_ubuntu
</details>

Description

When serving Qwen/Qwen3.5-35B-A3B-FP8 with MTP speculative decoding enabled, the model progressively produces incoherent tokens (random Unicode characters, repeated symbols, gibberish). The model is used in a RAG engine with text and images as inputs.

Launch command

vllm serve Qwen/Qwen3.5-35B-A3B-FP8 \
  --port x \
  --gpu-memory-utilization 0.95 \
  --max-model-len 62144 \
  --tensor-parallel-size 1 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}'

Observed behavior

During multi-turn chat completions with tool calling, the output quality degrades across consecutive requests. The speculative decoding metrics reflect this degradation: Request 1 (output is coherent):

SpecDecoding metrics: Mean acceptance length: 2.23, Avg Draft acceptance rate: 61.3%
Per-position acceptance rate: 0.798, 0.427

Request 2 (output starts to degrade):

SpecDecoding metrics: Mean acceptance length: 1.02, Avg Draft acceptance rate: 0.9%
Per-position acceptance rate: 0.018, 0.000

Request 3 (output is completely incoherent):

SpecDecoding metrics: Mean acceptance length: 1.00, Avg Draft acceptance rate: 0.0%
Per-position acceptance rate: 0.000, 0.000

The generated output contains random Unicode characters (e.g. ŧŧŧŧŧŧ), mixed-language fragments, and other garbage tokens.

Expected behavior

The draft acceptance rate should remain stable, or at least degrade gracefully without producing corrupted output.

Additional context

  • The AWQ-4bit variant (cyankiwi/Qwen3.5-35B-A3B-AWQ-4bit) with the same speculative config also exhibited repetition loops during generation.
  • After running some tests, disabling speculative decoding seems to resolve the issue, to confirm.

extent analysis

Fix Plan

To address the issue of incoherent tokens generated by the Qwen/Qwen3.5-35B-A3B-FP8 model with MTP speculative decoding enabled, we will attempt the following steps:

  1. Disable Speculative Decoding: As a temporary workaround, disable speculative decoding to confirm if it resolves the issue.
    • Modify the launch command to remove the --speculative-config flag or set "method" to null.
    • Example:

vllm serve Qwen/Qwen3.5-35B-A3B-FP8
--port x
--gpu-memory-utilization 0.95
--max-model-len 62144
--tensor-parallel-size 1
--reasoning-parser qwen3
--enable-auto-tool-choice
--tool-call-parser qwen3_coder

2. **Adjust Speculative Decoding Configuration**: If disabling speculative decoding is not feasible, try adjusting the speculative decoding configuration.
   - Reduce the `num_speculative_tokens` value to decrease the model's speculation.
   - Example:
     ```bash
vllm serve Qwen/Qwen3.5-35B-A3B-FP8 \
  --port x \
  --gpu-memory-utilization 0.95 \
  --max-model-len 62144 \
  --tensor-parallel-size 1 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":1}'
  1. Model Updates or Alternatives: Consider updating the model or exploring alternative models that may be less prone to speculation issues.
  2. Environment and Dependency Updates: Ensure that all dependencies, including PyTorch and CUDA, are up-to-date, as updates may include fixes for similar issues.

Verification

To verify the fix, monitor the model's output quality and speculative decoding metrics after applying the changes. Check for:

  • Coherent output without random Unicode characters or gibberish.
  • Stable draft acceptance rates.
  • Improved per-position acceptance rates.

Extra Tips

  • Regularly review and update models and dependencies to incorporate fixes and improvements.
  • Consider implementing additional logging or monitoring to detect and respond to output quality degradation.
  • Explore alternative speculative decoding methods or configurations that may better suit the specific use case.

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?
Loading…

FAQ

Expected behavior

The draft acceptance rate should remain stable, or at least degrade gracefully without producing corrupted output.

#api#ssr#installation#tensor shape#autograd error#model save/load#environment variable

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Back to top recommendations

TRENDING

vllm - ✅(Solved) Fix [Bug]: Gibberish output and collapsing generation throughput with Qwen3.5-35B-A3B-FP8 and speculative decoding enabled [1 pull requests, 1 participants]