Fix FP8 CUTLASS crash on SM12.1 (DGX Spark)

[ageev]

Summary

• Fixes "This kernel only supports sm120." crash when running FP8 models (e.g. Qwen3.5-35B-FP8) on DGX Spark (SM12.1)
• Replaces enable_sm120_only with enable_sm120_family in two CUTLASS kernel files, allowing SM12.1 (__CUDA_ARCH__ == 1210) to pass the architecture guard
• One-line sed applied at build time, before final compilation

This is a temporary workaround until the upstream fix is merged: vllm-project/vllm#35568

Details

vLLM's FP8 CUTLASS kernel wrappers use enable_sm120_only, which checks __CUDA_ARCH__ == 1200 and traps on any other arch. On SM12.1 (GB10, DGX Spark), __CUDA_ARCH__ is 1210, triggering cudaErrorLaunchFailure on every FP8 GEMM call.

enable_sm120_family already exists in the codebase (>= 1200 && < 1300) and is used by the blockwise dispatch path. This fix applies it to the two remaining files that still use enable_sm120_only.

See: https://github.com/saifgithub/vllm-gb10-sm121

Rebuild notes

Requires --rebuild-vllm since the fix is compiled into the binary. If upgrading from a previous build, clean Docker build cache first:

docker builder prune
./build-and-copy.sh --rebuild-vllm [--tf5] [-t vllm-node-tf5]

Note: --tf5 and -t vllm-node-tf5 are optional — the container name depends on your setup (default is vllm-node).

Test plan

• Verified on DGX Spark with Qwen3.5-35B-A3B-FP8 — model loads and serves without crash
• Confirmed enable_sm120_family symbols present in compiled _C.abi3.so