|
|
@@ -0,0 +1,48 @@
|
|
|
+# Qwen3.8-27B NVFP4 on 1x DGX Spark — the DGPP engine lane.
|
|
|
+#
|
|
|
+# 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) 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 peers itself. Stop
|
|
|
+# runs `scripts/dgpp-cluster down`. Readiness is /v1/models on :18080.
|
|
|
+#
|
|
|
+# Prereq: the DGPP repo checked out at `[dgpp].dir` on the launch node(s) with
|
|
|
+# `scripts/dgpp-cluster release` run once (builds the CUDA kernels). The engine
|
|
|
+# knobs below are DGPP's `engine` block, serialized verbatim into the JSON.
|
|
|
+
|
|
|
+engine = "dgpp"
|
|
|
+name = "Qwen3.8-27B NVFP4 (DGPP engine, 1x)"
|
|
|
+model = "unsloth/Qwen3.8-27B-NVFP4"
|
|
|
+parent_model = "Qwen/Qwen3.8-27B"
|
|
|
+port = 18080
|
|
|
+required_nodes = 1
|
|
|
+deploy = "native"
|
|
|
+
|
|
|
+[dgpp]
|
|
|
+http_port = 18080
|
|
|
+dir = "~/dgpp"
|
|
|
+
|
|
|
+# DGPP's engine block (from deploy/cluster_qwen3.8-27b_nvfp4_w1.example.json).
|
|
|
+[dgpp.engine]
|
|
|
+max_concurrency = 8
|
|
|
+kv_capacity = 262144
|
|
|
+kv_dtype = "bf16"
|
|
|
+default_max_tokens = 32768
|
|
|
+queue_limit = 8
|
|
|
+max_connections = 64
|
|
|
+decode_graph = true
|
|
|
+mtp = false
|
|
|
+dflash_model = "z-lab/Qwen3.8-27B-DFlash2"
|
|
|
+draft_head = "screen"
|
|
|
+dflash_weights = "fp8"
|
|
|
+prefill_activations = "native"
|
|
|
+prefix_cache_gib = 1.5
|
|
|
+admission = "grow"
|
|
|
+admission_window = 256
|
|
|
+model_alias = "Qwen3.8-27B-NVFP4"
|
|
|
+sampling_candidates = 128
|
|
|
+stats_interval_s = 10
|
|
|
+bf16_weights = "bf12"
|
