feat(recipe): deploy laguna-s2.1 via tmux with host venv (Docker image has wrong flashinfer)

recipes/laguna-s2.1-nvfp4/recipe.toml

@@ -14,6 +14,24 @@
 #
 # REFERENCE:
 #   https://huggingface.co/poolside/Laguna-S-2.1-NVFP4
+#
+# DEPLOYMENT: tmux (host venv). The pre-built Docker image
+# (ghcr.io/timothystewart6/vllm-gb10:v0.25.1-gb10.1) ships the wrong
+# flashinfer version, so Docker is non-functional. The sustainable path
+# is a host venv with the exact deps from the model card:
+#
+#   uv venv ~/venvs/vllm025 -p 3.12
+#   uv pip install vllm==0.25.1 --torch-backend=cu130
+#   uv pip install flashinfer-python==0.6.15.dev20260712 \
+#     flashinfer-cubin==0.6.15.dev20260712 \
+#     flashinfer-jit-cache==0.6.15.dev20260712 \
+#     --extra-index-url https://flashinfer.ai/whl/nightly/ \
+#     --extra-index-url https://flashinfer.ai/whl/nightly/cu130/
+#   hf download poolside/Laguna-S-2.1-NVFP4
+#   hf download poolside/Laguna-S-2.1-DFlash-NVFP4
+#
+# Then ensure `vllm` is on PATH (activate the venv or add its bin/ to PATH).
+# The entrypoint below (`["vllm", "serve"]`) is used directly by tmux.
 
 engine = "vllm"
 name = "Laguna S 2.1 NVFP4 (vLLM)"
@@ -21,6 +39,7 @@ docker_image = "ghcr.io/timothystewart6/vllm-gb10:v0.25.1-gb10.1"
 model = "poolside/Laguna-S-2.1-NVFP4"
 tensor_parallel_size = 1
 pipeline_parallel_size = 1
+deploy = "tmux"
 
 entrypoint = ["vllm", "serve"]
 
@@ -48,6 +67,7 @@ args = [
 SAFETENSORS_FAST_GPU = "1"
 TORCH_CUDA_ARCH_LIST = "12.1a"
 FLASHINFER_CUDA_ARCH_LIST = "12.1a"
+CUTE_DSL_ARCH = "sm_121a"
 VLLM_ALLOW_LONG_MAX_MODEL_LEN = "1"
 VLLM_FLOAT32_MATMUL_PRECISION = "high"
 PYTORCH_CUDA_ALLOC_CONF = "expandable_segments:True"