aiden lanes: entrypoint=[] for built-in-entrypoint images (launcher doubles CMD otherwise); vision 2x uses proven overlay img + aiden DSpark-greedy additions (his env-script needs a dedicated pass)

recipes/dsv4-vision-aiden-2x/recipe.toml

@@ -5,11 +5,14 @@
 # TP4/DCP1; on 2x expect roughly half). His production lane ships DCP2 + DSpark
 # (66-71 tok/s, 3.5M KV); DCP flags omitted here — DSpark runs fine under DCP1.
 #
-# Image: aidendle94/sparkrun-vllm-ds4-gb10:production-3.72-vision (published
-# vision stack: hand-ported tower/routing/CKV gather, cross-checked against
-# upstream vLLM PR #54566). Sibling comparison: dsv4-flash-vision:local (our
-# plugin overlay on his production-3.76.1 base) — the A/B pair for the vision
-# lane.
+# Image: dsv4-flash-vision:local (our plugin overlay on aiden's
+# production-3.76.1 base — see dsv4-flash-vision/ UPSTREAM.md). Aiden's
+# production-3.72-vision tag is the published 4x stack; running it under our
+# launcher needs his env-script contract (NODE_RANK/MASTER_ADDR/NCCL_IB_*
+# required by /usr/local/bin/deepseek-v4-flash-2x-spark-entrypoint →
+# /opt/vllm-src/serve-ds4-flash-spark.sh) — a dedicated integration pass,
+# tracked as follow-up. The overlay image carries the equivalent flags:
+# draft greedy + FULL_AND_PIECEWISE graphs + capture cap via the envs below.
 #
 # Weights: deepseek-ai/DeepSeek-V4-Flash-Vision-Exp (cached; ~84 GiB/node at
 # TP2 via ~/dsv4vision-hf NVMe copies). Draft heads ship in the checkpoint
@@ -28,14 +31,16 @@
 
 engine = "vllm"
 name = "DeepSeek V4 Flash Vision-Exp (aiden image, 2x Spark, DSpark-6)"
-docker_image = "aidendle94/sparkrun-vllm-ds4-gb10:production-3.72-vision"
+docker_image = "dsv4-flash-vision:local"
 model = "deepseek-ai/DeepSeek-V4-Flash-Vision-Exp"
 required_nodes = 2
 tensor_parallel_size = 2
 pipeline_parallel_size = 1
 distributed_backend = "mp"
 shm_size = "32gb"
-entrypoint = ["vllm", "serve"]
+# image ships ENTRYPOINT ["vllm","serve"]; our launcher passes the recipe
+# entrypoint as CMD words after the image, which would double it. Empty clears.
+entrypoint = []
 
 args = [
     "--served-model-name", "deepseek-v4-vision",

recipes/dsv4-vision-aiden-4x/recipe.toml

@@ -34,7 +34,9 @@ tensor_parallel_size = 4
 pipeline_parallel_size = 1
 distributed_backend = "mp"
 shm_size = "32gb"
-entrypoint = ["vllm", "serve"]
+# image ships ENTRYPOINT ["vllm","serve"]; our launcher passes the recipe
+# entrypoint as CMD words after the image, which would double it. Empty clears.
+entrypoint = []
 
 args = [
     "--served-model-name", "deepseek-v4-vision",

recipes/glm53-flash-aiden-2x/recipe.toml

@@ -39,7 +39,9 @@ tensor_parallel_size = 2
 pipeline_parallel_size = 1
 distributed_backend = "mp"
 shm_size = "32gb"
-entrypoint = ["vllm", "serve"]
+# image ships ENTRYPOINT ["vllm","serve"]; our launcher passes the recipe
+# entrypoint as CMD words after the image, which would double it. Empty clears.
+entrypoint = []
 
 args = [
     "--served-model-name", "glm53",

recipes/glm53-flash-aiden-4x/recipe.toml

@@ -31,7 +31,9 @@ tensor_parallel_size = 4
 pipeline_parallel_size = 1
 distributed_backend = "mp"
 shm_size = "32gb"
-entrypoint = ["vllm", "serve"]
+# image ships ENTRYPOINT ["vllm","serve"]; our launcher passes the recipe
+# entrypoint as CMD words after the image, which would double it. Empty clears.
+entrypoint = []
 
 args = [
     "--served-model-name", "glm53",