Upload nemotron-3-super.env with huggingface_hub

nemotron-3-super.env

@@ -0,0 +1,38 @@
+# NVIDIA Nemotron-3-Super-120B-A12B-NVFP4 — DGX Spark recipe
+# Source: huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4
+# 256k context, MTP speculative decoding ON (3 draft tokens).
+MODEL=nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4
+SERVED_NAME=nvidia/nemotron-3-super
+IMAGE=vllm/vllm-openai:v0.20.0
+GPU_MEM_UTIL=0.80
+# vllm/vllm-openai entrypoint already runs `vllm serve`.
+CMD_PREFIX=()
+
+EXTRA_DOCKER_ARGS=(
+  -e VLLM_NVFP4_GEMM_BACKEND=marlin
+  -e VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
+  -e VLLM_FLASHINFER_ALLREDUCE_BACKEND=trtllm
+  -e VLLM_USE_FLASHINFER_MOE_FP4=0
+  -v "$HOME/projects/serve/parsers/super_v3_reasoning_parser.py:/app/super_v3_reasoning_parser.py:ro"
+)
+
+EXTRA_VLLM_ARGS=(
+  --quantization fp4
+  --moe-backend marlin
+  --kv-cache-dtype fp8
+  --mamba_ssm_cache_dtype float16
+  --dtype auto
+  --tensor-parallel-size 1
+  --pipeline-parallel-size 1
+  --data-parallel-size 1
+  --max-model-len 262144
+  --max-num-seqs 4
+  --enable-chunked-prefill
+  --async-scheduling
+  --trust-remote-code
+  --reasoning-parser-plugin /app/super_v3_reasoning_parser.py
+  --reasoning-parser super_v3
+  --enable-auto-tool-choice
+  --tool-call-parser qwen3_coder
+  --speculative_config '{"method":"mtp","num_speculative_tokens":3,"moe_backend":"triton"}'
+)