[jamesETsmith]

Overview

The recipe marked all four AMD GPUs unsupported, because Beam's kernels were CUDA-only and only a CUDA image existed. vllm-project/vllm#62057 adds an AMD Beam implementation, validated on MI325X, MI350X, and MI355X, so this updates the recipe to match. Nothing changes for NVIDIA users.

🚧 leaving this PR as a draft until vllm-project/vllm#62057 is merged

Claims

• Beam runs on ROCm. MI325X (gfx942) and MI350X/MI355X (gfx950) are verified; MI300X is supported through the same validated gfx942 code path.
• BF16 only on AMD. So far only bf16 has been tested on AMD.
• BF16 fits every AMD node. 1204 GB fits MI300X (1536), MI325X (2048) and MI350X/MI355X (2304), so they join supported_hardware.
• DEP stays NVIDIA-only. The AMD path raises on sequence-parallel MoE, which is what DEP emits.
• Per-brand images. AMD points at vllm/vllm-openai-rocm:nightly, the image that will eventually carry #62057.
• AMD DFlash2 skips the attention backend pin. The NVIDIA drafter asserts FLASH_ATTN/FLASHINFER; the ROCm one delegates cache handling to the ROCm backend, so it doesn't need the pin.

Validation

• node scripts/build-recipes-api.mjs → ✓ JSON API: 215 models, no schema failures.

• node scripts/test-schema.mjs 4/4, node scripts/test-overlay.mjs 7/7.

• Checked the generated commands for all four AMD GPUs: BF16 variant, TP8, DEP disabled, and:

  vllm serve reflection/Beam-1.0-501B-A23B \
    --tensor-parallel-size 8 \
    --attention-backend ROCM_AITER_UNIFIED_ATTN \
    --moe-backend triton

• Serving evidence is #62057: BF16, TP8 + expert parallelism on MI325X, MI350X, and MI355X. Beam Verifier results reproduce Reflection's published BF16 accuracy within statistical error:

  • MI355X: GPQA-Diamond 90.06% and AIME 2026 97.45%, with both vanilla decoding and DFlash2.
  • MI325X: GPQA-Diamond 90.18%, with both vanilla decoding and DFlash2.

Details

One file, no schema or taxonomy changes. default_hardware stays b300; AMD users land on BF16 via default_variant. No NVIDIA behavior changed.

References

• [ROCm][Model] Add Beam 1.0 support on ROCm  vllm#62057 — Beam 1.0 on ROCm
• [Model] Add Reflection Beam vllm#61978 — day-0 Beam 1.0 (NVIDIA)
• Beam model card

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
vllm-recipes	Ready	Preview	Oct 23, 2026 6:22pm UTC

[gemini-code-assist]

Code Review

This pull request introduces AMD ROCm support for the Beam-1.0-501B-A23B model, enabling it to run on MI300X, MI325X, MI350X, and MI355X GPUs using the native BF16 checkpoint. It configures AMD-specific Docker images, hardware overrides, and attention/MoE backends, while updating the documentation guide to reflect these changes. The review feedback suggests enabling enable_adaptive_verification: true in both the AMD speculative decoding configuration and the corresponding guide example to maintain feature parity with the NVIDIA setup.

[gemini-code-assist]

The speculative decoding configuration for AMD ROCm is missing the enable_adaptive_verification: true option. Adaptive verification is a platform-agnostic vLLM feature that helps improve throughput under load by only verifying draft tokens likely to be accepted. It should be enabled on AMD as well to maintain feature parity with the NVIDIA configuration.

    hardware_overrides:\n      amd:\n        args:\n          - "--speculative-config"\n          - '{"method":"dflash","model":"reflection/Beam-1.0-501B-A23B-DFlash2","num_speculative_tokens":7,"enable_adaptive_verification":true}'

[gemini-code-assist]

Update the guide's speculative decoding example for AMD to include enable_adaptive_verification: true to match the recommended configuration.

  --speculative-config '{"method":"dflash","model":"reflection/Beam-1.0-501B-A23B-DFlash2","num_speculative_tokens":7,"enable_adaptive_verification":true}'

[jamesETsmith]

Missed #1107, so I'll add these recipes directly to my vllm PR vllm-project/vllm#62057