Home Explore Help
Sign In
sm121
/
launcher
2
0
Fork 0
Files Issues 0 Pull Requests 0 Wiki
Browse Source

kindling recipes: control paths are FILES, not the cache dirs — VLLM_{ADAPTIVE_K,DRAFT_TRUNC}_CONTROL pointed at the mounted dir roots, and their _read_control() does os.stat+open on the path (stat succeeds on a dir, open does not) => IsADirectoryError killed the engine core right after KV allocation; point them at <dir>/control (absent file = safe defaults: tau 0.3 / adapt, written file persists + live-control)

Alexander Korolev 3 weeks ago
parent
571a5499f1
commit
e6204c6b95
3 changed files with 6 additions and 6 deletions
Unified View Show Diff Stats
  1. 2 2
      recipes/glm53-kindling-2x/recipe.toml
  2. 2 2
      recipes/glm53-kindling-4x/recipe.toml
  3. 2 2
      recipes/glm53-kindling-6x/recipe.toml

+ 2 - 2
recipes/glm53-kindling-2x/recipe.toml
View File 

@@ -148,7 +148,7 @@ VLLM_ARXBIG_RS = "1"
 
 VLLM_ARXBIG_AG = "0"
 
 VLLM_ARXBIG_AG = "0"
 
 VLLM_ARX_TWO_SHOT_MAX_KB = "2048"
 
 VLLM_ARX_TWO_SHOT_MAX_KB = "2048"
 
 # adaptive-k.yaml — online step-cost learner + draft truncation
 
 # adaptive-k.yaml — online step-cost learner + draft truncation
 
-VLLM_ADAPTIVE_K_CONTROL = "/root/.cache/adaptive-k"
 
+VLLM_ADAPTIVE_K_CONTROL = "/root/.cache/adaptive-k/control"
 
 VLLM_ADAPTIVE_K_MODEL = "32.1,1.32,0.427"
 
 VLLM_ADAPTIVE_K_MODEL = "32.1,1.32,0.427"
 
 VLLM_ADAPTIVE_K_PER_REQUEST = "0"
 
 VLLM_ADAPTIVE_K_PER_REQUEST = "0"
 
 VLLM_ADAPTIVE_K_CODE_MODE = "gated"
 
 VLLM_ADAPTIVE_K_CODE_MODE = "gated"
 
@@ -156,7 +156,7 @@ VLLM_ADAPTIVE_K_ONLINE = "1"
 
 VLLM_ADAPTIVE_K_COST_MS = "2:29.7,3:33.1,4:36.5,5:39.8,6:43.1,7:46.0,8:48.9"
 
 VLLM_ADAPTIVE_K_COST_MS = "2:29.7,3:33.1,4:36.5,5:39.8,6:43.1,7:46.0,8:48.9"
 
 VLLM_DRAFT_TRUNC_TAU = "0.3"
 
 VLLM_DRAFT_TRUNC_TAU = "0.3"
 
 VLLM_DRAFT_TRUNC_EXPLORE = "4"
 
 VLLM_DRAFT_TRUNC_EXPLORE = "4"
 
-VLLM_DRAFT_TRUNC_CONTROL = "/root/.cache/draft-trunc"
 
+VLLM_DRAFT_TRUNC_CONTROL = "/root/.cache/draft-trunc/control"
 
 VLLM_DRAFT_TRUNC_LOG = ""
 
 VLLM_DRAFT_TRUNC_LOG = ""
 
 # fp8.yaml — dense linears the checkpoint kept in bf16 -> FP8/NVFP4 (the
 
 # fp8.yaml — dense linears the checkpoint kept in bf16 -> FP8/NVFP4 (the
 
 # measured ~1% NLL-on-prose cost is upstream's documented tuned default)
 
 # measured ~1% NLL-on-prose cost is upstream's documented tuned default)

+ 2 - 2
recipes/glm53-kindling-4x/recipe.toml
View File 

@@ -145,7 +145,7 @@ VLLM_ARXBIG_RS = "1"
 
 VLLM_ARXBIG_AG = "0"
 
 VLLM_ARXBIG_AG = "0"
 
 VLLM_ARX_TWO_SHOT_MAX_KB = "2048"
 
 VLLM_ARX_TWO_SHOT_MAX_KB = "2048"
 
 # adaptive-k.yaml — online step-cost learner + draft truncation
 
 # adaptive-k.yaml — online step-cost learner + draft truncation
 
-VLLM_ADAPTIVE_K_CONTROL = "/root/.cache/adaptive-k"
 
+VLLM_ADAPTIVE_K_CONTROL = "/root/.cache/adaptive-k/control"
 
 VLLM_ADAPTIVE_K_MODEL = "26.4,0.660,0.484"
 
 VLLM_ADAPTIVE_K_MODEL = "26.4,0.660,0.484"
 
 VLLM_ADAPTIVE_K_PER_REQUEST = "0"
 
 VLLM_ADAPTIVE_K_PER_REQUEST = "0"
 
 VLLM_ADAPTIVE_K_CODE_MODE = "gated"
 
 VLLM_ADAPTIVE_K_CODE_MODE = "gated"
 
@@ -153,7 +153,7 @@ VLLM_ADAPTIVE_K_ONLINE = "1"
 
 VLLM_ADAPTIVE_K_COST_MS = "2:29.7,3:33.1,4:36.5,5:39.8,6:43.1,7:46.0,8:48.9"
 
 VLLM_ADAPTIVE_K_COST_MS = "2:29.7,3:33.1,4:36.5,5:39.8,6:43.1,7:46.0,8:48.9"
 
 VLLM_DRAFT_TRUNC_TAU = "0.3"
 
 VLLM_DRAFT_TRUNC_TAU = "0.3"
 
 VLLM_DRAFT_TRUNC_EXPLORE = "4"
 
 VLLM_DRAFT_TRUNC_EXPLORE = "4"
 
-VLLM_DRAFT_TRUNC_CONTROL = "/root/.cache/draft-trunc"
 
+VLLM_DRAFT_TRUNC_CONTROL = "/root/.cache/draft-trunc/control"
 
 VLLM_DRAFT_TRUNC_LOG = ""
 
 VLLM_DRAFT_TRUNC_LOG = ""
 
 # fp8.yaml — dense linears the checkpoint kept in bf16 -> FP8/NVFP4 (the
 
 # fp8.yaml — dense linears the checkpoint kept in bf16 -> FP8/NVFP4 (the
 
 # measured ~1% NLL-on-prose cost is upstream's documented tuned default)
 
 # measured ~1% NLL-on-prose cost is upstream's documented tuned default)

+ 2 - 2
recipes/glm53-kindling-6x/recipe.toml
View File 

@@ -150,7 +150,7 @@ VLLM_ARXBIG_RS = "1"
 
 VLLM_ARXBIG_AG = "0"
 
 VLLM_ARXBIG_AG = "0"
 
 VLLM_ARX_TWO_SHOT_MAX_KB = "2048"
 
 VLLM_ARX_TWO_SHOT_MAX_KB = "2048"
 
 # adaptive-k.yaml — online step-cost learner + draft truncation
 
 # adaptive-k.yaml — online step-cost learner + draft truncation
 
-VLLM_ADAPTIVE_K_CONTROL = "/root/.cache/adaptive-k"
 
+VLLM_ADAPTIVE_K_CONTROL = "/root/.cache/adaptive-k/control"
 
 VLLM_ADAPTIVE_K_MODEL = "19.3,0.50,0.40"
 
 VLLM_ADAPTIVE_K_MODEL = "19.3,0.50,0.40"
 
 VLLM_ADAPTIVE_K_PER_REQUEST = "0"
 
 VLLM_ADAPTIVE_K_PER_REQUEST = "0"
 
 VLLM_ADAPTIVE_K_CODE_MODE = "gated"
 
 VLLM_ADAPTIVE_K_CODE_MODE = "gated"
 
@@ -158,7 +158,7 @@ VLLM_ADAPTIVE_K_ONLINE = "1"
 
 VLLM_ADAPTIVE_K_COST_MS = "2:29.7,3:33.1,4:36.5,5:39.8,6:43.1,7:46.0,8:48.9"
 
 VLLM_ADAPTIVE_K_COST_MS = "2:29.7,3:33.1,4:36.5,5:39.8,6:43.1,7:46.0,8:48.9"
 
 VLLM_DRAFT_TRUNC_TAU = "0.3"
 
 VLLM_DRAFT_TRUNC_TAU = "0.3"
 
 VLLM_DRAFT_TRUNC_EXPLORE = "4"
 
 VLLM_DRAFT_TRUNC_EXPLORE = "4"
 
-VLLM_DRAFT_TRUNC_CONTROL = "/root/.cache/draft-trunc"
 
+VLLM_DRAFT_TRUNC_CONTROL = "/root/.cache/draft-trunc/control"
 
 VLLM_DRAFT_TRUNC_LOG = ""
 
 VLLM_DRAFT_TRUNC_LOG = ""
 
 # fp8.yaml — dense linears the checkpoint kept in bf16 -> FP8/NVFP4 (the
 
 # fp8.yaml — dense linears the checkpoint kept in bf16 -> FP8/NVFP4 (the
 
 # measured ~1% NLL-on-prose cost is upstream's documented tuned default)
 
 # measured ~1% NLL-on-prose cost is upstream's documented tuned default)