Add DeepSeek V4 Flash Recipe

[SpaceshipCreative]

These results are impressive, and this recipe should be added.

https://forums.developer.nvidia.com/t/deepseek-v4-flash-aiden-recipe-from-reddit-1m-token-session-operational-cuda-12-1-tailored-for-dgx-spark-gb10/372268

[VinceDu66]

https://forums.developer.nvidia.com/t/deepseek-v4-flash-aiden-recipe-from-reddit-1m-token-session-operational-cuda-12-1-tailored-for-dgx-spark-gb10/372268/112

A little bit modified with "gpu_memory_utilization: 0.15" , "max_num_seqs: 6" and "--kv-cache-memory-bytes 5.6g"

...It's a shame it's not multimodal; Claude Code is very fond of that ;)

recipe_version: "1"
name: deepseek-v4-flash
description: DeepSeek V4 Flash FP8 on dual DGX Spark TP=2
model: deepseek-ai/DeepSeek-V4-Flash
container: vllm-node-dsv4
cluster_only: true
mods: []
defaults:
  port: 8000
  host: 0.0.0.0
  tensor_parallel: 2
  pipeline_parallel: 1
  gpu_memory_utilization: 0.15
  block_size: 256
  max_model_len: 500000
  max_num_batched_tokens: 8192
  max_num_seqs: 6
  served_model_name: deepseek-v4-flash
env:
  HF_HOME: /root/.cache/huggingface
  HF_HUB_OFFLINE: "1"
  TRANSFORMERS_OFFLINE: "1"
  VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
  VLLM_USE_B12X_MOE: "1"
  VLLM_SPARSE_INDEXER_MAX_LOGITS_MB: "256"
  VLLM_TRITON_MLA_SPARSE: "1"
  TORCH_CUDA_ARCH_LIST: 12.1a
  TRITON_CACHE_DIR: "/root/.triton"
  VLLM_ENGINE_READY_TIMEOUT_S: "7200"
command: |
  vllm serve deepseek-ai/DeepSeek-V4-Flash \
      --served-model-name {served_model_name} \
      --host {host} \
      --port {port} \
      --trust-remote-code \
      --tensor-parallel-size {tensor_parallel} \
      --pipeline-parallel-size {pipeline_parallel} \
      --kv-cache-memory-bytes 5.6g \
      --kv-cache-dtype fp8 \
      --block-size {block_size} \
      --enable-prefix-caching \
      --max-model-len {max_model_len} \
      --max-num-seqs {max_num_seqs} \
      --enable-chunked-prefill \
      --max-num-batched-tokens {max_num_batched_tokens} \
      --gpu-memory-utilization {gpu_memory_utilization} \
      --distributed-executor-backend mp \
      --speculative-config '{{"method":"mtp","num_speculative_tokens":2}}' \
      --tokenizer-mode deepseek_v4 \
      --tool-call-parser deepseek_v4 \
      --enable-auto-tool-choice \
      --reasoning-parser deepseek_v4 \
      --reasoning-config '{{"reasoning_parser":"deepseek_v4","reasoning_start_str":"<think>","reasoning_end_str":"</think>"}}' \
      --default-chat-template-kwargs '{{"thinking":true,"preserve_thinking":true}}' \
      --load-format safetensors \
      --nnodes 2