prep recipes: ling-31-flash-nvfp4 (4x, BailingMoeV3 on v0.31.0, weights pending) + step5-preview-nvfp4 (8x, dedicated step-5-preview image, VL, GB10-unverified)

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

@@ -0,0 +1,29 @@
+# Ling-3.1-flash (inclusionAI) on 4x DGX Spark — PREP recipe (weights pending).
+#
+# 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.
+#
+# 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.
+
+engine = "vllm"
+name = "Ling-3.1-flash NVFP4 (4x, PREP — weights pending)"
+docker_image = "vllm/vllm-openai:v0.31.0"
+model = "cbfight/Ling-3.1-flash-NVFP4-lossless8"
+parent_model = "inclusionAI/Ling-3.1-flash"
+port = 30000
+required_nodes = 4
+tensor_parallel_size = 4
+args = [
+    "--trust-remote-code",
+    "--max-model-len", "65536",
+    "--gpu-memory-utilization", "0.78",
+]

recipes/step5-preview-nvfp4/recipe.toml

@@ -0,0 +1,28 @@
+# Step-5-Preview-NVFP4 (stepfun-ai) on 8x DGX Spark — PREP recipe (GB10 probe).
+#
+# Status 2026-10-16: `Step5ForConditionalGeneration` (`model_type: step5_vl`) —
+# a VISION-LANGUAGE model (8-bit, modelopt/NVFP4), 604B, ~473 GB. vLLM v0.31.0
+# registers `StepVLForConditionalGeneration` (step_vl) but NOT
+# `Step5ForConditionalGeneration` (step5_vl) — a newer arch — so stock v0.31.0
+# won't load it. Use the DEDICATED image `vllm/vllm-openai:step-5-preview`
+# (prebuilt external, no docker_image_dir). 8 Sparks (TP8).
+#
+# ⚠ NOT verified on DGX Spark / GB10 (sm_121a): expect kernel/arch gaps and the
+# unified-memory gate (vLLM #48140 reads MemFree; large max-model-len may refuse
+# to boot). This is a feasibility probe, not a bench lane. Needs the 8x cluster
+# up + a ~473 GB download — run in a supervised window. See
+# specs/tech-architecture/upstream-watch-2026-10-16.md.
+
+engine = "vllm"
+name = "Step-5-Preview NVFP4 (8x, PREP — GB10 unverified)"
+docker_image = "vllm/vllm-openai:step-5-preview"
+model = "stepfun-ai/Step-5-Preview-NVFP4"
+parent_model = "stepfun-ai/Step-5-Preview"
+port = 30000
+required_nodes = 8
+tensor_parallel_size = 8
+args = [
+    "--trust-remote-code",
+    "--max-model-len", "32768",
+    "--gpu-memory-utilization", "0.78",
+]