docker-compose.eval.yml

@@ -25,3 +25,58 @@ services:
             - driver: nvidia
               count: 1
               capabilities: [gpu]
+
+  # Fase 5: clona 1:1 el docker-compose.yml real de produccion (vllm-qwen36, ver
+  # PLAN.md) -- mismos flags de vLLM (incluido --speculative-config real, la
+  # primera vez que se prueba en este proyecto), cambiando solo container_name,
+  # puerto (8002, produccion usa 8000 y el vllm-eval de Fase 4 usa 8001), volumen
+  # (checkpoint NVFP4 de Fase 5 en vez de RedHatAI--Qwen3.6-35B-A3B-NVFP4),
+  # --model/--served-model-name, y restart: "no". Nunca se toca vllm-qwen36 ni su
+  # compose real de Portainer -- este es un contenedor nuevo y propio del repo.
+  vllm-eval-nvfp4:
+    image: vllm/vllm-openai:cu130-nightly-aarch64
+    container_name: vllm-eval-nvfp4
+    restart: "no"
+    runtime: nvidia
+    environment:
+      NVIDIA_VISIBLE_DEVICES: all
+      NVIDIA_DRIVER_CAPABILITIES: compute,utility
+    ports:
+      - "8002:8000"
+    ipc: host
+    ulimits:
+      memlock: -1
+      stack: 67108864
+    volumes:
+      - /home/aleleba/ft-models/Qwen3.6-35B-A3B-mcp-NVFP4:/models/Qwen3.6-35B-A3B-mcp-NVFP4:ro
+    command: >
+      --model /models/Qwen3.6-35B-A3B-mcp-NVFP4
+      --served-model-name qwen3.6-35b-a3b-mcp-nvfp4
+      --host 0.0.0.0
+      --port 8000
+      --tensor-parallel-size 1
+      --trust-remote-code
+      --quantization compressed-tensors
+      --moe-backend flashinfer_cutlass
+      --kv-cache-dtype fp8_e4m3
+      --gpu-memory-utilization 0.45
+      --max-model-len 524288
+      --max-num-seqs 8
+      --max-num-batched-tokens 32768
+      --enable-chunked-prefill
+      --enable-prefix-caching
+      --speculative-config '{"method":"mtp","num_speculative_tokens":1}'
+      --reasoning-parser qwen3
+      --tool-call-parser qwen3_coder
+      --enable-auto-tool-choice
+      --default-chat-template-kwargs '{"preserve_thinking":true}'
+      --limit-mm-per-prompt '{"image":4,"video":0,"audio":0}'
+      --generation-config vllm
+      --override-generation-config '{"temperature":0.6,"top_p":0.80,"top_k":20,"presence_penalty":0.0,"repetition_penalty":1.0}'
+      --hf-overrides '{"text_config":{"rope_scaling":{"rope_type":"yarn","factor":2.0,"original_max_position_embeddings":262144}}}'
+    healthcheck:
+      test: ["CMD", "curl", "-f", "http://localhost:8000/health"]
+      interval: 30s
+      timeout: 10s
+      retries: 5
+      start_period: 600s