[Support Matrix] Fill DGDR coverage gaps for DeepSeek-R1, Nemotron-3 Super, and Qwen3-32B

[sttts]

What coverage are you requesting?

The DGDR/Sweeper comparison suite exposes four reproducible AIConfigurator performance-data gaps. Legal, KV-feasible configurations are found, but Replay cannot evaluate any sampled candidate for the affected path.

Model	Hardware	Backend / version	Mode	Missing data
deepseek-ai/DeepSeek-R1	h200_sxm	SGLang 0.5.14	disaggregated	Context MLA for fmha_quant=fp8, kv_quant=fp8
nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8	h200_sxm	SGLang 0.5.14	aggregated and disaggregated	FP8 MoE data for the model's 512-expert shape
nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8	h200_sxm	vLLM 0.24.0	aggregated	FP8 MoE data for the model's 512-expert shape
Qwen/Qwen3-32B-FP8	a100_sxm	vLLM 0.14.0	disaggregated	fp8_block GEMM data; only BF16 is available

These correspond to six cases because DeepSeek is tested with two GPU budgets and Nemotron/SGLang in both deployment modes.

Reproduction and actual errors

The failures reproduce with AIConfigurator and AIConfigurator Core 0.11.0, using the native RustForwardPassPerfModel through Dynamo Replay:

DeepSeek-R1 / H200 / SGLang:
context MLA data missing for ContextKey { fmha_quant: "fp8", kv_quant: "fp8" }

Nemotron-3 Super / H200 / SGLang:
MoE data missing for MoeKey { quant: "fp8", distribution: "uniform",
topk: 22, num_experts: 512, hidden_size: 1024, inter_size: 2688,
moe_tp_size: 1, moe_ep_size: 2 }

Nemotron-3 Super / H200 / vLLM:
MoE data missing for MoeKey { quant: "fp8", distribution: "power_law_1.01",
topk: 22, num_experts: 512, hidden_size: 1024, inter_size: 2688,
moe_tp_size: 1, moe_ep_size: 2 }

Qwen3-32B-FP8 / A100 / vLLM:
GEMM perf data missing for quant 'fp8_block'; available: ["bfloat16"]

Expected behavior

Either:

1. package performance data that covers these configurations; or
2. reject them during a readiness/preflight check with a typed unsupported-coverage result, before a Sweeper search repeatedly evaluates candidates that cannot be predicted.

Additional context

• #850 registered DeepSeek-R1 but did not add the missing H200/SGLang FP8 Context-MLA data.
• #1522 fixed the analogous Nemotron FP8 MoE gap for H100/vLLM; H200 vLLM and SGLang remain affected.
• #1414 covers fp8-block data quality, but not the absence of A100/vLLM rows reported here.
• The originating cases are in ai-dynamo/dynamo#13765, based on ai-dynamo/dynamo#8469.

[sttts]

Related defects found by the same 29-case comparison: ai-dynamo/aisimulate#123 tracks the single-GPU MoE enumeration bug, ai-dynamo/aisimulate#124 tracks the misleading terminal viability error, and ai-dynamo/dynamo#14423 tracks missing TRT-LLM/disaggregated Replay support.