From d1e98929115c3866c14db28da38ffa471ee0332a Mon Sep 17 00:00:00 2001
From: Alejandro Lembke Barrientos <aleleba@hotmail.com>
Date: Thu, 30 Jul 2026 01:10:36 +0000
Subject: [PATCH] Fase 5: servicio de diagnostico vllm-eval-nvfp4-nospec
 (puerto 8003)

Identico a vllm-eval-nvfp4 pero sin --speculative-config, para aislar si la
regresion de calidad observada en las puertas 2-3 (vs. Fase 4) viene del
speculative decoding (MTP) o de la cuantizacion NVFP4 en si -- decision
explicita del usuario antes de invertir tiempo en recalibrar con mas
muestras de calibracion.
---
 docker-compose.eval.yml | 53 +++++++++++++++++++++++++++++++++++++++++
 1 file changed, 53 insertions(+)

diff --git a/docker-compose.eval.yml b/docker-compose.eval.yml
index e969ddb..6f092c9 100644
--- a/docker-compose.eval.yml
+++ b/docker-compose.eval.yml
@@ -80,3 +80,56 @@ services:
       timeout: 10s
       retries: 5
       start_period: 600s
+
+  # Fase 5 -- diagnostico de aislamiento: identico a vllm-eval-nvfp4 pero SIN
+  # --speculative-config, para determinar si la regresion de calidad observada
+  # en las puertas 2-3 (vs. Fase 4) viene del speculative decoding (MTP) o de
+  # la cuantizacion NVFP4 en si. Puerto 8003 (distinto de 8000 produccion, 8001
+  # Fase 4, 8002 Fase 5 con speculative real). Servicio temporal de diagnostico,
+  # no clona produccion 1:1 a proposito (esa es la variable que se esta aislando).
+  vllm-eval-nvfp4-nospec:
+    image: vllm/vllm-openai:cu130-nightly-aarch64
+    container_name: vllm-eval-nvfp4-nospec
+    restart: "no"
+    runtime: nvidia
+    environment:
+      NVIDIA_VISIBLE_DEVICES: all
+      NVIDIA_DRIVER_CAPABILITIES: compute,utility
+    ports:
+      - "8003:8000"
+    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-nospec
+      --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
+      --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