feat: DeepSeek V4 Flash DSpark C12 NVFP4 recipe + mod

[mtdphn]

spark-vllm-docker v1 style recipe for DeepSeek-V4-Flash-DSpark

Eugr-format recipe for serving DeepSeek-V4-Flash-DSpark on 2x DGX Spark (GB10) with vLLM TP=2, DSpark speculative decoding, NVFP4 KV cache (nvfp4_ds_mla), B12X MoE backend, and the 2026-07-03 garble fix.

Includes

• Recipe: recipes/deepseek-v4-flash-dspark.yaml — v1 format, cluster-only (2-node)
• Mod: mods/deepseek-v4-flash-dspark/ — 18 overlay files + 3 NVFP4 patch stages

Key Parameters

Param	Value
container	vllm-dspark-runtime:dspark-nvfp4-stage-c
gpu_memory_utilization	0.77
max_model_len	350000
kv_cache_dtype	nvfp4_ds_mla
max_num_seqs	12
num_speculative_tokens	5
distributed-executor-backend	mp

Attribution

• tonyd2wild — C12 NVFP4 default config (canonical 350K verified profile)
• MiaAI-Lab — DSpark concurrency patch (Keys C12)
• 0rand, renek — DSpark cold-start garble root cause analysis
• Aiden Le — production-3.2 image reference, fp8 fallback guidance

Verified

• Deployed on 2x DGX Spark (back-to-back, 100G fabric)
• Tool-eval-bench: 87/100 (15 scenarios, 10 trials)
• 13 passed, 1 partial, 1 failed
• Median turn time: 2.4s

Notes

• Tested on Gigabyte Atom heads at gpu_memory_utilization=0.77 (~94 GiB free). Clean Spark units can use 0.80.
• The mod's run.sh verification script previously emitted bare print(...) calls (missing f-string prefix) causing benign SyntaxWarning during import verification — fixed in this submission.