|
|
@@ -16,17 +16,24 @@
|
|
|
# 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
|
|
|
+# (ghcr.io/timothystewart6/vllm-gb10:v0.26.0-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:
|
|
|
+# is a host venv with the exact deps from the model card.
|
|
|
+#
|
|
|
+# ACTIVE (risky): vLLM 0.26.0 + flashinfer 0.6.16.dev20260729.
|
|
|
+# 0.6.16 adds SM121 (GB10/DGX Spark) NVFP4 attention support
|
|
|
+# (flashinfer PR #3897) and Expert Parallelism, but introduces
|
|
|
+# API changes that vLLM 0.26.0 (which requires flashinfer 0.6.14)
|
|
|
+# may not fully track. Tested on dgx1.lan — monitor for MoE cubin
|
|
|
+# validation regressions ("reject incompatible output-scale cubins").
|
|
|
+#
|
|
|
+# SAFE (commented out): vLLM 0.25.1 + flashinfer 0.6.15.dev20260728.
|
|
|
+# Same minor as the original 0.6.15.dev20260712 — just 16 days of
|
|
|
+# bug fixes. No API changes. Drop-in replacement.
|
|
|
#
|
|
|
# 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/
|
|
|
+# uv pip install flashinfer-python==0.6.15.dev20260728 flashinfer-cubin==0.6.15.dev20260728 flashinfer-jit-cache==0.6.15.dev20260728 --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
|
|
|
#
|
|
|
@@ -35,16 +42,16 @@
|
|
|
|
|
|
engine = "vllm"
|
|
|
name = "Laguna S 2.1 NVFP4 (vLLM)"
|
|
|
-docker_image = "ghcr.io/timothystewart6/vllm-gb10:v0.25.1-gb10.1"
|
|
|
+docker_image = "ghcr.io/timothystewart6/vllm-gb10:v0.26.0-gb10.1"
|
|
|
model = "poolside/Laguna-S-2.1-NVFP4"
|
|
|
tensor_parallel_size = 1
|
|
|
pipeline_parallel_size = 1
|
|
|
deploy = "tmux"
|
|
|
-venv = "~/venvs/vllm025"
|
|
|
+venv = "~/venvs/vllm026"
|
|
|
venv_setup = """
|
|
|
- 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/
|
|
|
+ uv venv ~/venvs/vllm026 -p 3.12
|
|
|
+ uv pip install vllm==0.26.0 --torch-backend=cu130
|
|
|
+ uv pip install flashinfer-python==0.6.16.dev20260729 flashinfer-cubin==0.6.16.dev20260729 flashinfer-jit-cache==0.6.16.dev20260729 --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
|
|
|
"""
|
