|
|
@@ -38,9 +38,12 @@ tensor_parallel_size = 2
|
|
|
pipeline_parallel_size = 1
|
|
|
distributed_backend = "mp"
|
|
|
shm_size = "32gb"
|
|
|
-# image ships ENTRYPOINT ["vllm","serve"]; our launcher passes the recipe
|
|
|
-# entrypoint as CMD words after the image, which would double it. Empty clears.
|
|
|
-entrypoint = []
|
|
|
+# his image has TWO vllm wheels: dist-packages (plain, NO vision) and the
|
|
|
+# compiled stack at /opt/vllm-src/.venv (multimodal-wrapped DeepseekV4). The
|
|
|
+# image's own entrypoint script also injects a big env block (B12X_*,
|
|
|
+# PCIE_ALLREDUCE, MEGA_AOT, JIT paths) — the flag-relevant subset is in [env]
|
|
|
+# below. Point the launcher tail straight at their venv python.
|
|
|
+entrypoint = ["/opt/vllm-src/.venv/bin/python", "-m", "vllm.entrypoints.cli.main", "serve"]
|
|
|
|
|
|
args = [
|
|
|
"--served-model-name", "deepseek-v4-vision",
|
|
|
@@ -81,6 +84,14 @@ args = [
|
|
|
[env]
|
|
|
# aidee runner envs (his guide, minus DCP2-only and fabric knobs)
|
|
|
CUTE_DSL_ARCH = "sm_121a"
|
|
|
+TRITON_PTXAS_PATH = "/usr/local/cuda/bin/ptxas"
|
|
|
+DG_JIT_USE_NVRTC = "0"
|
|
|
+USE_CUDNN = "1"
|
|
|
+VLLM_USE_MEGA_AOT_ARTIFACT = "1"
|
|
|
+VLLM_ENABLE_PCIE_ALLREDUCE = "1"
|
|
|
+VLLM_PCIE_ALLREDUCE_BACKEND = "b12x"
|
|
|
+VLLM_USE_B12X_DENSE_SPLITK_TURBO = "1"
|
|
|
+VLLM_USE_B12X_W4A16_TC_DECODE = "1"
|
|
|
VLLM_ENABLE_DSPARK = "1"
|
|
|
VLLM_ENABLE_MTP = "0"
|
|
|
NUM_SPECULATIVE_TOKENS = "6"
|
