[Bug]: `dynamic_func() got multiple values for argument 'HEAD_DIM'` crash during startup when using `-dcp`

[josephrocca]

Your current environment

The output of python collect_env.py

/root/venv/lib/python3.12/site-packages/torch/cuda/__init__.py:63: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
  import pynvml  # type: ignore[import]
Collecting environment information...
uv is set
==============================
        System Info
==============================
OS                           : Ubuntu 24.04.3 LTS (x86_64)
GCC version                  : (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
Clang version                : Could not collect
CMake version                : Could not collect
Libc version                 : glibc-2.39

==============================
       PyTorch Info
==============================
PyTorch version              : 2.8.0+cu128
Is debug build               : False
CUDA used to build PyTorch   : 12.8
ROCM used to build PyTorch   : N/A

==============================
      Python Environment
==============================
Python version               : 3.12.3 (main, Aug 14 2025, 17:47:21) [GCC 13.3.0] (64-bit runtime)
Python platform              : Linux-6.8.0-79-generic-x86_64-with-glibc2.39

==============================
       CUDA / GPU Info
==============================
Is CUDA available            : True
CUDA runtime version         : 12.8.93
CUDA_MODULE_LOADING set to   : LAZY
GPU models and configuration : 
GPU 0: NVIDIA H200
GPU 1: NVIDIA H200
GPU 2: NVIDIA H200
GPU 3: NVIDIA H200
GPU 4: NVIDIA H200
GPU 5: NVIDIA H200
GPU 6: NVIDIA H200
GPU 7: NVIDIA H200

Nvidia driver version        : 580.82.07
cuDNN version                : Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.9.12.0
/usr/lib/x86_64-linux-gnu/libcudnn_adv.so.9.12.0
/usr/lib/x86_64-linux-gnu/libcudnn_cnn.so.9.12.0
/usr/lib/x86_64-linux-gnu/libcudnn_engines_precompiled.so.9.12.0
/usr/lib/x86_64-linux-gnu/libcudnn_engines_runtime_compiled.so.9.12.0
/usr/lib/x86_64-linux-gnu/libcudnn_graph.so.9.12.0
/usr/lib/x86_64-linux-gnu/libcudnn_heuristic.so.9.12.0
/usr/lib/x86_64-linux-gnu/libcudnn_ops.so.9.12.0
HIP runtime version          : N/A
MIOpen runtime version       : N/A
Is XNNPACK available         : True

==============================
          CPU Info
==============================
Architecture:                         x86_64
CPU op-mode(s):                       32-bit, 64-bit
Address sizes:                        52 bits physical, 57 bits virtual
Byte Order:                           Little Endian
CPU(s):                               176
On-line CPU(s) list:                  0-175
Vendor ID:                            AuthenticAMD
BIOS Vendor ID:                       QEMU
Model name:                           AMD EPYC 9654 96-Core Processor
BIOS Model name:                      pc-q35-8.2  CPU @ 2.0GHz
BIOS CPU family:                      1
CPU family:                           25
Model:                                17
Thread(s) per core:                   1
Core(s) per socket:                   1
Socket(s):                            176
Stepping:                             1
BogoMIPS:                             4799.99
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 fsrm arch_capabilities
Virtualization:                       AMD-V
Hypervisor vendor:                    KVM
Virtualization type:                  full
L1d cache:                            11 MiB (176 instances)
L1i cache:                            11 MiB (176 instances)
L2 cache:                             88 MiB (176 instances)
L3 cache:                             2.8 GiB (176 instances)
NUMA node(s):                         1
NUMA node0 CPU(s):                    0-175
Vulnerability Gather data sampling:   Not affected
Vulnerability Itlb multihit:          Not affected
Vulnerability L1tf:                   Not affected
Vulnerability Mds:                    Not affected
Vulnerability Meltdown:               Not affected
Vulnerability Mmio stale data:        Not affected
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed:               Not affected
Vulnerability Spec rstack overflow:   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 Tsx async abort:        Not affected

==============================
Versions of relevant libraries
==============================
[pip3] flashinfer-python==0.3.1
[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==1.14.1
[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-ml-py==13.580.82
[pip3] nvidia-nccl-cu12==2.27.3
[pip3] nvidia-nvjitlink-cu12==12.8.93
[pip3] nvidia-nvtx-cu12==12.8.90
[pip3] pynvml==13.0.1
[pip3] pyzmq==27.1.0
[pip3] torch==2.8.0+cu128
[pip3] torchaudio==2.8.0+cu128
[pip3] torchvision==0.23.0+cu128
[pip3] transformers==4.56.2
[pip3] triton==3.4.0
[conda] Could not collect

==============================
         vLLM Info
==============================
ROCM Version                 : Could not collect
vLLM Version                 : 0.11.0rc2.dev110+g1405f0c7b (git sha: 1405f0c7b)
vLLM Build Flags:
  CUDA Archs: Not Set; ROCm: Disabled
GPU Topology:
        GPU0    GPU1    GPU2    GPU3    GPU4    GPU5    GPU6    GPU7    CPU Affinity    NUMA Affinity   GPU NUMA ID
GPU0     X      NV18    NV18    NV18    NV18    NV18    NV18    NV18    0-175   0               N/A
GPU1    NV18     X      NV18    NV18    NV18    NV18    NV18    NV18    0-175   0               N/A
GPU2    NV18    NV18     X      NV18    NV18    NV18    NV18    NV18    0-175   0               N/A
GPU3    NV18    NV18    NV18     X      NV18    NV18    NV18    NV18    0-175   0               N/A
GPU4    NV18    NV18    NV18    NV18     X      NV18    NV18    NV18    0-175   0               N/A
GPU5    NV18    NV18    NV18    NV18    NV18     X      NV18    NV18    0-175   0               N/A
GPU6    NV18    NV18    NV18    NV18    NV18    NV18     X      NV18    0-175   0               N/A
GPU7    NV18    NV18    NV18    NV18    NV18    NV18    NV18     X      0-175   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
CUDA_MODULE_LOADING=LAZY

🐛 Describe the bug

vLLM crashes with dynamic_func() got multiple values for argument 'HEAD_DIM' during startup when using -dcp 4:

# ~latest commit as of writing
export VLLM_COMMIT=1405f0c7bab17b10682d2de337bd544b3e9f2f92 && uv pip install vllm[flashinfer] --torch-backend=auto --extra-index-url https://wheels.vllm.ai/${VLLM_COMMIT}
CUDA_VISIBLE_DEVICES='0,1,2,3' vllm serve RedHatAI/DeepSeek-R1-0528-quantized.w4a16 --port 3000 --tensor-parallel-size 4 -dcp 4 --served-model-name default --host 0.0.0.0

The issue does not exist in the latest stable version, 0.10.2.

Startup / crash logs:

https://gist.github.com/josephrocca/af3ae259dd63c5392245131fa58e7d7f

Note: For the log lines like Worker_TP0 .... and Worker_TP1 ...., etc. I've removed all but the TP0 lines, since they were of course duplicated 4 times.

[josephrocca]

These two edits (from Claude) seem to fix the bug:

  ⎿  Updated vllm/vllm/attention/ops/common.py with 23 additions and 2 removals            
       120                         cp_group: GroupCoordinator,
       121                         ctx: CPTritonContext = None):
       122        """
       123 -      cp_attn_out: [ B, H, D ]
       124 -      cp_attn_lse: [ B, H ]
       123 +      cp_attn_out: [ B, H, D ] or [ B, 1, H, D ] (decode with seq_len=1)
       124 +      cp_attn_lse: [ B, H ] or [ B, H, 1 ] (decode with seq_len=1)
       125        """
       126        if cp_group.world_size == 1:
       127            return cp_attn_out
     ...
       129        if ctx is None:
       130            ctx = CPTritonContext()
       131    
       132 +      # Handle 4D tensors from decode (with seq_len=1 dimension)
       133 +      # Squeeze to 3D for the kernel which expects [B, H, D] and [N, B, H]
       134 +      original_out_shape = cp_attn_out.shape
       135 +      if cp_attn_out.ndim == 4:
       136 +          # Shape is [B, 1, H, D], squeeze to [B, H, D]
       137 +          assert cp_attn_out.shape[1] == 1, \
       138 +              f"Expected seq_len=1, got {cp_attn_out.shape[1]}"
       139 +          cp_attn_out = cp_attn_out.squeeze(1)
       140 +  
       141 +      if cp_attn_lse.ndim == 3:
       142 +          # Shape is [B, H, 1], squeeze to [B, H]
       143 +          assert cp_attn_lse.shape[-1] == 1, \
       144 +              f"Expected seq_len=1, got {cp_attn_lse.shape[-1]}"
       145 +          cp_attn_lse = cp_attn_lse.squeeze(-1)
       146 +  
       147        lses = torch.empty((cp_group.world_size, ) + cp_attn_lse.shape,
       148                           dtype=cp_attn_lse.dtype,
       149                           device=cp_attn_lse.device)
     ...
       137        lses = cp_group.all_gather(cp_attn_lse, dim=0).view_as(lses)
       138        out, _ = correct_attn_out(cp_attn_out, lses, cp_group.rank_in_group, ctx)
       139        out = cp_group.reduce_scatter(out, dim=1)
       140 +  
       141 +      # Restore original shape if needed
       142 +      if len(original_out_shape) == 4:
       143 +          # Restore [B, 1, H, D] shape
       144 +          out = out.unsqueeze(1)
       145 +  
       146        return out

  ⎿  Updated vllm/vllm/v1/engine/core.py with 6 additions                                          
       157                    or self.scheduler.get_kv_connector() is not None):
       158    
       159                block_size = vllm_config.cache_config.block_size
       160 +              # Use logical block size for hashing when DCP is enabled.
       161 +              # With DCP, KV cache is interleaved across ranks, so the logical
       162 +              # block size is block_size * dcp_world_size.
       163 +              dcp_world_size = vllm_config.parallel_config.decode_context_parallel_size
       164 +              if dcp_world_size > 1:
       165 +                  block_size *= dcp_world_size
       166                caching_hash_fn = get_hash_fn_by_name(
       167                    vllm_config.cache_config.prefix_caching_hash_algo)
       168                init_none_hash(caching_hash_fn)

Without the latter edit, the KV prefix cache of requests get "mixed up" - requests basically retrieve the wrong context tokens from the cache IIUC.

You can reproduce the fix like this (tested on 4xH200):

cd /root
git clone https://github.com/vllm-project/vllm && cd vllm && git reset --hard decf7f794bfff2dd591f79e591c2e5292fe9fe0f
wget -O "/root/vllm/vllm/attention/ops/common.py" "https://gist.githubusercontent.com/josephrocca/f52de175a9dd7e10c6f80253f0bcebc2/raw/228ae302a2dbfce79837d4db161f25e6d53704eb/common.py"
wget -O "/root/vllm/vllm/v1/engine/core.py" "https://gist.githubusercontent.com/josephrocca/f52de175a9dd7e10c6f80253f0bcebc2/raw/228ae302a2dbfce79837d4db161f25e6d53704eb/core.py"
VLLM_USE_PRECOMPILED=1 uv pip install --editable .[flashinfer]
VLLM_USE_FLASHINFER_SAMPLER=0 CUDA_VISIBLE_DEVICES='0,1,2,3' vllm serve RedHatAI/DeepSeek-R1-0528-quantized.w4a16 --tensor-parallel-size 4 -dcp 4 --served-model-name default --max-num-seqs 512 --max-model-len 9216

CC @youzhedian @LucasWilkinson

I'm still getting output quality issues, but it's more subtle, and I think that might be related to a different issue, specifically this one:

• [Bug]: TypeError: argument 'id': StreamInput must be either an integer or a list of integers #25821

[josephrocca]

I'm still getting output quality issues, but it's more subtle, and I think that might be related to a different issue

Turns out this was due to needing to explicitly set cuda graph mode to 'PIECEWISE', since 'FULL' is not supported by DCP yet, but there is no guard against this in the code.

So the above two edits, plus --compilation-config '{"cudagraph_mode":"PIECEWISE"}' fixes the issue.

[josephrocca]

HEAD_DIM error solved by:

• CP: make correct_attn_out robust to 4‑D views and fix Triton arg binding #26509

Gibberish error is a separate issue, and as mentioned above is solved by adding --compilation-config '{"cudagraph_mode":"PIECEWISE"}' when using DCP.