vulcan-decision-shell-v1-runtime

中文

Ready-to-load ONNX packages of the three vulcan-decision-shell-v1 tool-call judges, for automatic download by VulcanCode and the dsh plugin. The three models differ only in size: every package has the same files, the same graph interface and the same config schema.

Model	Parameters	Pick on CUDA and WebGPU	Pick on CPU
mini	0.2B	mini/fp16 (487 MB)	mini/fp32 (800 MB)
flash	0.36B	flash/fp16 (822 MB)	flash/fp16
standard	0.8B	standard/fp16 (1.53 GB)	standard/fp16

catalog.json holds the same choice for programs (families.<model>.recommended.<device>), the manifest digest that pins each package, and the measured speed and limits of every format.

Package layout

MANIFEST.json                     every other file with its size and SHA-256
runtime-config.json               limits, decision margin, graph interface, text encoding
model.onnx                        input_ids, attention_mask [batch, sequence] int64 -> logits [batch, 3] float32
assets/policy-v2.json             policy text the weights were trained with
assets/tokenizer.json             tokenizer of this model
example-input.json                nine self-check calls, the same in every package
reports/delivery-validation.json  expected decision and token count of each self-check call

Logits are in allow, review, block order. Rows are right-padded; attention_mask marks the real tokens. The highest score wins, except that allow must lead the best other score by allow_margin (1.0 for mini, 0 for the others) or the call becomes review. Two things are fixed by training and declared in runtime-config.json under encoding: mini reads a policy head with three [MASK] markers, flash and standard read one English decision prompt.

Deployment parameters

Every package's runtime-config.json holds these values and catalog.json repeats them under families.<model>.runtime; all formats of one model share them. A host may submit many calls in one request: it sorts them by length and closes a batch once it holds micro_batch rows or once rows × longest row would exceed padded_batch_tokens. A single call longer than max_tokens is not evaluated.

Parameter	mini	flash	standard
max_tokens (per call)	8192	32768	32768
max_request_items (calls per request)	256	256	256
micro_batch (rows per batch)	8	16	32
padded_batch_tokens (rows × longest row per batch)	8192	8192	8192
allow_margin	1.0	0	0
gpu_memory_limit_mib	12288	12288	12288

Requirements and limits

• ONNX Runtime 1.30.0 (CUDA 13 build for GPU). Tested on RTX 5080 16 GB.
• max_tokens is 32768 for flash and standard (a 32K call takes 1.5 s and 2.8 s on GPU) and 8192 for mini; longer calls are not evaluated.
• mini stops at 8192 because its base model context is 8192 and longer calls are unproven.
• CPU works but is slow: about 0.1 s (mini), 0.2 s (flash) and 0.35 s (standard) per short call.
• WebGPU (GPUs without CUDA) runs every package through the ONNX Runtime WebGPU plugin: in batches of ten, about 40 ms (mini), 17 ms (flash) and 31 ms (standard) per call. Tested on the RTX 5080 only.

Other precisions (BF16, INT8, full FP16) stay in the per-model ONNX repositories; they change decisions or bring no speed gain, which catalog.json records under not_included.