DeepEP / expert-parallel NVSHMEM kernels unavailable on GB10: arch list hardcodes 9.0a/10.0a

[timothystewart6]

Summary

DeepEP expert-parallel NVSHMEM kernels are not available on GB10 (sm_121a). The kernel build system hardcodes 9.0a (Hopper) and 10.0a (datacenter Blackwell) in its arch list, excluding sm_121a.

Impact

• Expert parallelism via DeepEP is unavailable
• Standard tensor-parallel inference is unaffected
• Attempting to use DeepEP kernels will fail at runtime

What is needed upstream

DeepEP needs to add sm_121a to its NVSHMEM kernel arch list and validate on GB10 hardware.

vLLM commit

Verified against v0.23.1rc0 (e3e3cd54589cee689b785aab5bda81b3e4203191).

[timothystewart6]

Upstream context to track before rebuilding DeepEP / expert-parallel paths for GB10

As of 2026-06-25, this issue was verified against this repo's current vLLM base:

• vLLM v0.23.1rc0
• vLLM commit e3e3cd54589cee689b785aab5bda81b3e4203191

Concrete upstream refs found:

1. DeepEP #167, merged

   • PR: Make TORCH_CUDA_ARCH_LIST as an environment variable deepseek-ai/DeepEP#167
   • Title: Make TORCH_CUDA_ARCH_LIST as an environment variable
   • Merge commit: d5ca4495c0c068bc617102841b9322d378fac8ea
   • Merged: 2025-05-19
   • The PR body says it adds 10.0 to TORCH_CUDA_ARCH_LIST for Blackwell compilation.
   • This is related Blackwell build plumbing, but it is not an explicit sm_121a / GB10 enablement PR.

2. DeepEP #168, upstream issue

   • Issue: Support compilation for Blackwell by adding "10.0" to the TORCH_CUDA_ARCH_LIST deepseek-ai/DeepEP#168
   • Title: Support compilation for Blackwell by adding "10.0" to the TORCH_CUDA_ARCH_LIST
   • This is the issue associated with adding Blackwell architecture support for 10.0, not 12.1a.

3. DeepEP #519, upstream issue

   • Issue: assert not compatible on Blackwell GPUs deepseek-ai/DeepEP#519
   • Title: assert not compatible on Blackwell GPUs
   • Related to Blackwell build/runtime compatibility errors.

4. DeepEP #550, open

   • PR: feat: add auto-detection for Blackwell GPU architecture (sm_100a) deepseek-ai/DeepEP#550
   • Title: feat: add auto-detection for Blackwell GPU architecture (sm_100a)
   • The PR body says it adds automatic GPU architecture detection for Blackwell sm_100a / 10.0a and Hopper sm_90a.
   • It is open and not merged as of this comment.
   • It is related Blackwell build plumbing, but it does not explicitly add sm_121a / GB10 support.

5. DeepEP #341, merged on hybrid-ep

   • PR: Support nvfp4 low latency mode dispatch deepseek-ai/DeepEP#341
   • Title: Support nvfp4 low latency mode dispatch
   • Merge commit: ef73fd9a4c098e09929e48dfd696505ddc8e2043
   • Merged: 2025-10-23 into hybrid-ep.
   • Related to NVFP4 low-latency dispatch, but not an sm_121a architecture-list fix.

Search result note:

• I searched DeepEP upstream for sm120, sm121, GB10, DGX Spark, Blackwell, 12.1, sm_121, and 121a.
• I did not find an upstream DeepEP PR or issue that explicitly adds sm_121a / GB10 to the NVSHMEM kernel arch list.
• The concrete upstream refs above are related Blackwell / DeepEP build or runtime items, but none are an exact resolution for this issue's sm_121a arch-list gap.

Build-tracking note:

• Standard tensor-parallel inference remains separate from this DeepEP expert-parallel path.
• Before enabling DeepEP / expert parallel on GB10, track whether DeepEP upstream adds explicit sm_121a / compute capability 12.1 support in its build and NVSHMEM kernel path, or whether vLLM carries a separate GB10-specific DeepEP integration.