Add DSv4 Flash recipe for DGX Station

[prodney-dev]

Also add a per hardware default strategy for overriding the overall default strategy

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
vllm-recipes	Ready	Preview	Jul 31, 2026 7:38pm

[gemini-code-assist]

Code Review

This pull request introduces support for per-hardware default strategy recommendations (via default_strategy_hardware) and applies it to the DeepSeek-V4-Flash model on the DGX Station GB300 hardware. It also updates the DGX Station single-GPU serving guide with optimized vLLM parameters. A critical shadowing bug was identified in the review where the exact-GPU-ID overrides (such as dgx_station_gb300) are overwritten by generation-level overrides (blackwell) due to the evaluation order in command synthesis.

[prodney-dev]

@esmeetu , can you review this one?

[faradawn]

The configs looks good, e.g. moe backend = auto.

1. Instead of changing the hardware ID to override, how about we add a markdown section under dsv4 flash that instruct users to use TP on dgx station. Then, you can also mention why.
2. is the claude.md changes neccessry?

[prodney-dev]

The configs looks good, e.g. moe backend = auto.

1. Instead of changing the hardware ID to override, how about we add a markdown section under dsv4 flash that instruct users to use TP on dgx station. Then, you can also mention why.

If we don't change the default recipe, then it will be broken on DGX Station. I don't think that will be a good experience. We need at least a functional recipe at the top level and more optimization can be added in the markdown section

2. is the claude.md changes neccessry?

Yes, this tells coding agents about new fields/features which are added.

[esmeetu]

It could be simple if just use and don't need touch js code:

strategy_overrides:
  single_node_tp:
    extra_args:
      - "--moe-backend"
      - "auto"

TP should use auto moe backend instead of deep_gemm_mega_moe.

[prodney-dev]

It could be simple if just use and don't need touch js code:

strategy_overrides:
  single_node_tp:
    extra_args:
      - "--moe-backend"
      - "auto"

TP should use auto moe backend instead of deep_gemm_mega_moe.

The idea of default_strategy_hardware is to add a per hardware default strategy, so whenever a user selects a hardware, the default strategy specific to that hardware should be selected and the default recipe that they see should work on that hardware. In this case the default strategy was single_node_tep but we want to change it to single_node_tp for DGX station.

[prodney-dev]

@esmeetu, let me know if you have any more concerns or we can merge this PR?

[esmeetu]

@prodney-dev Can we make default strategy to single_node_tp, and don't need to introduce other complexity.

[prodney-dev]

@prodney-dev Can we make default strategy to single_node_tp, and don't need to introduce other complexity.

@esmeetu, I can change that but I haven't verified on other hardware. If it is okay, I can make the change.