glm53-flash-nvfp4-fp8-dgpp-2x: DGPP-engine GLM-5.3-Flash w2 recipe (engine=dgpp, deploy=native, world_size=2, [dgpp.engine] from DGPP w2 example)

recipes/glm53-flash-nvfp4-fp8-dgpp-2x/recipe.toml

@@ -0,0 +1,49 @@
+# GLM-5.3-Flash NVFP4+FP8 on 2x DGX Spark — the DGPP engine lane (w2).
+#
+# DGPP (github.com/HawkBearPig/dgpp) is a from-scratch C++/CUDA inference
+# engine for DGX Spark. First-class `engine = "dgpp"` + `deploy = "native"`:
+# the launcher generates DGPP's cluster JSON (model + world_size + engine
+# knobs + http port + paths) and a sibling `.env` (node list + ssh user +
+# ports) from this recipe + the selected cluster, uploads both to rank 0's
+# DGPP checkout (`[dgpp].dir`, default ~/dgpp), and runs
+# `scripts/dgpp-cluster up --config <path>` there. Head-coordinated: the
+# manager SSHes to the peer itself. Stop runs `scripts/dgpp-cluster down`.
+# Readiness is /v1/models on :18080.
+#
+# world_size = 2 (required_nodes = 2): DGPP splits the model across two Sparks
+# and the rank-0 manager brings up rank 1. The engine knobs below are DGPP's
+# `engine` block from deploy/cluster_glm-5.3-flash_nvfp4-fp8_w2.example.json,
+# serialized verbatim into the generated JSON.
+#
+# Prereq: the DGPP repo checked out at `[dgpp].dir` on BOTH launch nodes with
+# `scripts/dgpp-cluster release` run once (builds the CUDA kernels).
+
+engine = "dgpp"
+name = "GLM-5.3-Flash NVFP4+FP8 (DGPP engine, 2x)"
+model = "HawkBearPig/GLM-5.3-Flash-NVFP4-FP8"
+parent_model = "zai-org/GLM-5.3-Flash"
+port = 18080
+required_nodes = 2
+deploy = "native"
+
+[dgpp]
+http_port = 18080
+dir = "~/dgpp"
+
+# DGPP's engine block (from deploy/cluster_glm-5.3-flash_nvfp4-fp8_w2.example.json).
+[dgpp.engine]
+max_concurrency = 4
+kv_capacity = 163840
+kv_dtype = "fp8"
+bf16_weights = "bf12"
+default_max_tokens = 32768
+queue_limit = 8
+max_connections = 64
+decode_graph = true
+mtp = true
+draft_head = "screen"
+prefix_cache_gib = 4.5
+admission = "full"
+admission_window = 256
+sampling_candidates = 128
+stats_interval_s = 10