feat: add deepseek-v4-flash-dspark recipe

- New recipe for DeepSeek-V4-Flash-DSpark with DSpark speculative decoding
- Image: aidendle94/sparkrun-vllm-ds4-gb10:production-3.6
- Uses dspark method (not mtp), b12x MoE, async scheduling
- Model: deepseek-ai/DeepSeek-V4-Flash-DSpark
- Port 8000, shm_size 32gb, --served-model-name ChatGPTN

recipes/deepseek-v4-flash-dspark/recipe.toml

@@ -0,0 +1,98 @@
+# DeepSeek-V4-Flash-DSpark — 2× GB10 / DGX Spark, TP=2 over RoCE.
+# Uses the DSpark speculative decoding method with aidendle94's production-3.6 image.
+#
+# Image: aidendle94/sparkrun-vllm-ds4-gb10:production-3.6
+#   Latest GB10-native image with DSpark support, b12x MoE, and multi-node fixes.
+# Model: deepseek-ai/DeepSeek-V4-Flash-DSpark
+#   The DSpark variant tuned for speculative decoding on DGX Spark.
+#
+# Reference: Nvidia Developer Forum — DeepSeek-V4-Flash-DSpark recipe
+
+engine = "vllm"
+name = "DeepSeek V4 Flash DSpark (vLLM)"
+docker_image = "aidendle94/sparkrun-vllm-ds4-gb10:production-3.6"
+model = "deepseek-ai/DeepSeek-V4-Flash-DSpark"
+port = 8000
+required_nodes = 2
+tensor_parallel_size = 2
+pipeline_parallel_size = 1
+distributed_backend = "mp"
+shm_size = "32gb"
+
+entrypoint = ["vllm", "serve"]
+
+args = [
+    "--kv-cache-dtype", "fp8",
+    "--block-size", "256",
+    "--max-model-len", "393216",
+    "--max-num-seqs", "128",
+    "--max-num-batched-tokens", "4096",
+    "--max-cudagraph-capture-size", "128",
+    "--gpu-memory-utilization", "0.8",
+    "--served-model-name", "ChatGPTN",
+    # Engine backends
+    "--attention-backend", "FLASHINFER_MLA_SPARSE_DSV4",
+    "--moe-backend", "b12x",
+    # Chunked prefill + prefix caching
+    "--enable-chunked-prefill",
+    "--enable-prefix-caching",
+    "--enable-flashinfer-autotune",
+    # Compilation
+    "--compilation-config", """
+'{"cudagraph_mode":"FULL_AND_PIECEWISE","custom_ops":["all"]}'""",
+    # Async scheduling
+    "--async-scheduling",
+    "--no-scheduler-reserve-full-isl",
+    # DSpark speculative decoding
+    "--speculative-config", """
+'{"method":"dspark","num_speculative_tokens":3,"draft_sample_method":"probabilistic"}'""",
+    # Tokenizer / reasoning
+    "--tokenizer-mode", "deepseek_v4",
+    "--tool-call-parser", "deepseek_v4",
+    "--reasoning-parser", "deepseek_v4",
+    "--enable-auto-tool-choice",
+    "--default-chat-template-kwargs.thinking=true",
+    "--default-chat-template-kwargs.reasoning_effort=max",
+    # Disable custom all-reduce (not needed on GB10 single-NIC)
+    "--disable-custom-all-reduce",
+    "--trust-remote-code",
+]
+
+[env]
+# DSpark speculative decoding
+VLLM_DSPARK_REPLICATE_MARKOV_W1 = "1"
+# b12x MoE (this image's primary path)
+VLLM_USE_B12X_MOE = "1"
+
+# b12x suite = REQUIRED by this image
+VLLM_USE_B12X_MHC = "1"
+VLLM_USE_B12X_WO_PROJECTION = "1"
+VLLM_USE_B12X_FP8_GEMM = "1"
+VLLM_USE_B12X_SPARSE_INDEXER = "1"
+VLLM_USE_V2_MODEL_RUNNER = "1"
+VLLM_USE_FLASHINFER_SAMPLER = "1"
+VLLM_USE_AOT_COMPILE = "1"
+VLLM_USE_MEGA_AOT_ARTIFACT = "0"
+VLLM_USE_BREAKABLE_CUDAGRAPH = "0"
+VLLM_MEMORY_PROFILE_INCLUDE_ATTN = "1"
+VLLM_PREFIX_CACHE_RETENTION_INTERVAL = "4096"
+B12X_MLA_SM120_UNIFIED = "1"
+B12X_DENSE_SPLITK_TURBO = "1"
+B12X_W4A16_TC_DECODE = "1"
+B12X_MHC_MAX_TOKENS = "16384"
+
+# GPU memory management
+CUTE_DSL_ARCH = "sm_121a"
+PYTORCH_CUDA_ALLOC_CONF = "expandable_segments:True"
+OMP_NUM_THREADS = "4"
+
+# Offline mode
+HF_HUB_OFFLINE = "1"
+
+# Compile-time arch hints
+TORCH_CUDA_ARCH_LIST = "12.1a"
+
+# NCCL overrides (cluster config ib_hca/socket_ifname used for most; recipe wins last)
+NCCL_P2P_LEVEL = "SYS"
+NCCL_CUMEM_ENABLE = "0"
+