|
|
@@ -135,12 +135,24 @@ container = "/root/.cache/torchinductor"
|
|
|
host = "~/sm121-cache/jit-121f/vllm"
|
|
|
container = "/root/.cache/vllm"
|
|
|
|
|
|
+# Triton's own JIT cache (the vllm compile cache references cubins that live
|
|
|
+# here + in TORCHINDUCTOR_CACHE_DIR; without these, every restart re-JITs and
|
|
|
+# TritonBundler logs "cubin not found").
|
|
|
+[[extra_mounts]]
|
|
|
+host = "~/sm121-cache/jit-121f/triton"
|
|
|
+container = "/root/.cache/triton"
|
|
|
+
|
|
|
[env]
|
|
|
HF_HUB_OFFLINE = "1"
|
|
|
VLLM_PLUGINS = "b12x_loader"
|
|
|
VLLM_ENABLE_ROCE_ALLREDUCE = "1"
|
|
|
MAX_JOBS = "2"
|
|
|
VLLM_ENGINE_READY_TIMEOUT_S = "3600"
|
|
|
+# Keep inductor + triton caches on the mounted volumes (defaults are the
|
|
|
+# container /tmp and ~/.triton — both die every restart, leaving the vllm
|
|
|
+# compile cache dangling at lost cubins).
|
|
|
+TORCHINDUCTOR_CACHE_DIR = "/root/.cache/torchinductor"
|
|
|
+TRITON_CACHE_DIR = "/root/.cache/triton"
|
|
|
# SparkRing tp4 serving-profile keys (quality flags from the DS4.1 garble
|
|
|
# line + b12x policy + v2 runner); NCCL transport stays on the switched
|
|
|
# cluster config — their switchless-ring keys are intentionally not ported.
|
