[Support Matrix] Request coverage for DeepSeek-V4-Flash on RTX PRO 6000D with SGLang 0.5.16

[Ray-ling7]

What coverage are you requesting?

Model (HuggingFace ID):

deepseek-ai/DeepSeek-V4-Flash

Native FP4-expert checkpoint, architecture DeepseekV4ForCausalLM.

Hardware / System:

• Product designation: NVIDIA RTX PRO 6000D
• CUDA/NVML-reported device name: NVIDIA RTX 6000D
• Compute capability: SM120
• Runtime-visible memory: 83.05 GiB per GPU
• AIC system definition: not yet upstreamed; naming to be confirmed

This is the Chinese-market RTX PRO 6000D version and should not be
treated as equivalent to the existing 96 GB rtx_pro_6000_server
system definition.

Backend / Framework:

SGLang

Backend Version (if specific):

SGLang 0.5.16

Mode

• Aggregated (agg)
• Disaggregated (disagg)
• Both

Additional context

The existing DeepSeek-V4 SGLang collectors contain several API
assumptions specific to SGLang 0.5.14 and cannot run against SGLang
0.5.16 without compatibility updates.

The affected collection paths include:

• DeepSeek-V4 CSA/HCA context and generation modules
• MHC pre/post modules
• MXFP4 MoE
• FP8 block GEMM

Representative incompatibilities observed with SGLang 0.5.16:

• ModelRunner.__init__() no longer accepts the legacy tp_rank
  arguments
• Req.set_extend_input_len() is no longer available
• ForwardBatch.init_new() requires
  return_hidden_states_before_norm
• KV-cache profiling moved from ModelRunner into
  KVCacheConfigurator
• the legacy _global_server_args access path is no longer available
• MoE and FP8 kernel modules were relocated

I have implemented a local compatibility patch and validated the
collectors against the environment above.

Local validation completed:

• 213 relevant collector tests passed
• Ruff lint passed
• Ruff formatting check passed
• CSA and HCA context collection completed successfully
• CSA and HCA generation collection completed successfully
• MHC, MoE, and FP8 block GEMM probes completed successfully
• repeated measurements were stable

The implementation can be split into collector-focused PRs:

1. DeepSeek-V4 CSA/HCA context and generation compatibility
2. MHC collector compatibility
3. MXFP4 MoE collector compatibility
4. SM120 FP8 block GEMM collector compatibility

I would also like to contribute the validated SGLang 0.5.16 collector
compatibility changes and unit tests in a follow-up PR.