feat: force colored vLLM logs via VLLM_LOGGING_COLOR=1

vLLM's default 'auto' mode disables ANSI colors when stdout is not
a terminal (e.g. piped through Docker).  Forcing to "1" preserves
colors in 'cargo run logs --recipe ... --follow' output.

recipes/deepseek-v4-flash/recipe.toml

@@ -48,6 +48,9 @@ RAY_memory_monitor_refresh_ms = "0"
 RAY_num_prestart_python_workers = "0"
 RAY_object_store_memory = "1073741824"
 
+# Force colored logs through Docker pipeline
+VLLM_LOGGING_COLOR = "1"
+
 # vLLM features for DeepSeek V4 on Blackwell (GB10 / sm_120)
 VLLM_ALLOW_LONG_MAX_MODEL_LEN = "1"
 VLLM_USE_B12X_MOE = "1"

recipes/minimax-m3-mxfp8/recipe.toml

@@ -1,6 +1,6 @@
 engine = "vllm"
 name = "MiniMax M3 MXFP8 (vLLM)"
-docker_image = "aidendle94/sparkrun-vllm-ds4-gb10:production-v2"
+docker_image = "vllm/vllm-openai:minimax-m3"
 model = "MiniMaxAI/MiniMax-M3-MXFP8"
 required_nodes = 4
 tensor_parallel_size = 4
@@ -12,12 +12,14 @@ distributed_backend = "mp"
 # MSA sparse attention requires --block-size 128.
 # Text-only mode skips the vision encoder to save VRAM.
 
-entrypoint = ["vllm", "serve"]
+#entrypoint = ["vllm", "serve"]
 
 args = [
     "--block-size", "128",
     "--kv-cache-dtype", "fp8",
     "--gpu-memory-utilization", "0.85",
+    "--load-format", "safetensors",
+    "--enable-prefix-caching",
     "--tool-call-parser", "minimax_m3",
     "--reasoning-parser", "minimax_m3",
     "--enable-auto-tool-choice",
@@ -28,5 +30,11 @@ args = [
 ]
 
 [env]
+# Force colored logs through Docker pipeline
+VLLM_LOGGING_COLOR = "1"
+
 # M3 supports thinking via chat_template_kwargs.thinking_mode
 # enabled / adaptive / disabled — pass per-request in extra_body
+SAFETENSORS_FAST_GPU = "1"
+VLLM_ALLOW_LONG_MAX_MODEL_LEN = "1"
+VLLM_USE_B12X_MOE = "1"