ling/step5 recipes: point at the downloaded weights (Ling fp8 original 535G, Step-5 NVFP4 368G on NAS) + entrypoint=[vllm,serve] (build_tail needs it or the model lands in --entrypoint); Ling 8x TP8 stock v0.31.0

recipes/ling-31-flash-nvfp4/recipe.toml

@@ -1,29 +1,38 @@
-# Ling-3.1-flash (inclusionAI) on 4x DGX Spark — PREP recipe (weights pending).
+# Ling-3.1-flash (inclusionAI) fp8 on 8x DGX Spark.
 #
-# Status 2026-10-16: the arch is supported on our base but the small weights do
-# not exist yet. Ling-3.1-flash is `BailingMoeV3ForCausalLM` (bailing_hybrid,
-# custom_code, fp8) — vLLM v0.31.0 registers `BailingMoeV3ForCausalLM` AND
-# `BailingMoeV3MTPModel` (MTP), so NO nightly is needed. The day-0 NVFP4 cut
-# `cbfight/Ling-3.1-flash-NVFP4-lossless8` is an EMPTY repo (reserved, not
-# populated), so this 4x lane targets that cut once it lands. The only
-# loadable Ling today is the 572B fp8 original (~689 GB → 8 Sparks); see
-# specs/tech-architecture/upstream-watch-2026-10-16.md.
+# 2026-10-16: the fp8 checkpoint is on the NAS HF cache (535 GB, snapshot
+# a4ccd6b4). Arch `BailingMoeV3ForCausalLM` (+ `BailingMoeV3MTPModel`) is
+# registered in stock vLLM v0.31.0 — no nightly, no custom image. 535 GB / 8 =
+# ~67 GB weights/rank, fits a GB10 with KV headroom.
 #
 # Base: stock `vllm/vllm-openai:v0.31.0` (prebuilt, arch built-in) — NOT the
-# kindling image (its overlays are GLM-5.3-specific). `--trust-remote-code` is
-# required (the model ships custom_code auto_map). Flip `model` to the real
-# NVFP4 repo when cbfight populates it; keep TP=4 for the ~GLM-5.3-Flash class.
+# kindling image (its overlays are GLM-5.3-specific). `--trust-remote-code` for
+# the model's custom_code auto_map (vLLM still prefers its native bailing_moe_v3
+# impl). max-model-len kept modest (65536) to stay clear of the GB10 unified-
+# memory boot gate (vLLM #48140 reads MemFree; 128K+ can refuse to boot) — raise
+# once it's proven up. GMU 0.78 (the measured GB10 safe ceiling).
+#
+# GB10/sm_121a fp8-MoE feasibility is UNVERIFIED for this arch — first live
+# probe. See specs/tech-architecture/upstream-watch-2026-10-16.md.
 
 engine = "vllm"
-name = "Ling-3.1-flash NVFP4 (4x, PREP — weights pending)"
+name = "Ling-3.1-flash fp8 (8x, stock v0.31.0)"
 docker_image = "vllm/vllm-openai:v0.31.0"
-model = "cbfight/Ling-3.1-flash-NVFP4-lossless8"
+entrypoint = ["vllm", "serve"]
+model = "inclusionAI/Ling-3.1-flash"
 parent_model = "inclusionAI/Ling-3.1-flash"
 port = 30000
-required_nodes = 4
-tensor_parallel_size = 4
+required_nodes = 8
+tensor_parallel_size = 8
+pipeline_parallel_size = 1
+distributed_backend = "mp"
 args = [
     "--trust-remote-code",
     "--max-model-len", "65536",
     "--gpu-memory-utilization", "0.78",
+    "--max-num-seqs", "32",
+    "--max-num-batched-tokens", "8192",
 ]
+
+[env]
+VLLM_ENGINE_READY_TIMEOUT_S = "3600"

recipes/step5-preview-nvfp4/recipe.toml

@@ -16,6 +16,7 @@
 engine = "vllm"
 name = "Step-5-Preview NVFP4 (8x, PREP — GB10 unverified)"
 docker_image = "vllm/vllm-openai:step-5-preview"
+entrypoint = ["vllm", "serve"]
 model = "stepfun-ai/Step-5-Preview-NVFP4"
 parent_model = "stepfun-ai/Step-5-Preview"
 port = 30000