mirror of
https://github.com/open-jarvis/OpenJarvis.git
synced 2026-08-14 00:47:52 +00:00
Consolidates NVIDIA vLLM, Apple Silicon, CPU Pearl mining support, CLI/docs, and live H100 validation.
99 lines
3.0 KiB
YAML
99 lines
3.0 KiB
YAML
name: Pearl Model Validation
|
|
description: Track conversion and validation of a Pearl-compatible mining model
|
|
labels: ["type:feature", "area:mining"]
|
|
body:
|
|
- type: markdown
|
|
attributes:
|
|
value: |
|
|
Use this template when promoting a raw Hugging Face model to a
|
|
Pearl-compatible `pearl-ai/*-pearl` mining model. A model should remain
|
|
`planned` in OpenJarvis until this checklist is complete.
|
|
- type: input
|
|
id: raw_model
|
|
attributes:
|
|
label: Raw model
|
|
placeholder: e.g., Qwen/Qwen3.5-9B
|
|
validations:
|
|
required: true
|
|
- type: input
|
|
id: pearl_model
|
|
attributes:
|
|
label: Pearl model artifact
|
|
placeholder: e.g., pearl-ai/Qwen3.5-9B-pearl
|
|
validations:
|
|
required: true
|
|
- type: dropdown
|
|
id: target_provider
|
|
attributes:
|
|
label: Target provider
|
|
options:
|
|
- vllm-pearl
|
|
- cpu-pearl
|
|
- apple-mps-pearl
|
|
validations:
|
|
required: true
|
|
- type: textarea
|
|
id: quantization_recipe
|
|
attributes:
|
|
label: Quantization recipe
|
|
description: Link or paste the recipe used to create the Pearl model artifact.
|
|
placeholder: |
|
|
- compressed-tensors config:
|
|
- 7-bit mining layers:
|
|
- 8-bit non-mining layers:
|
|
- calibration data:
|
|
- SmoothQuant settings:
|
|
validations:
|
|
required: true
|
|
- type: textarea
|
|
id: hardware
|
|
attributes:
|
|
label: Validation hardware
|
|
placeholder: |
|
|
- GPU:
|
|
- VRAM:
|
|
- driver/CUDA:
|
|
- Docker image/tag:
|
|
- Pearl commit/ref:
|
|
validations:
|
|
required: true
|
|
- type: checkboxes
|
|
id: acceptance
|
|
attributes:
|
|
label: Acceptance checks
|
|
options:
|
|
- label: Model loads in Pearl's vLLM miner container
|
|
required: true
|
|
- label: vLLM registers Pearl's quantization plugin
|
|
required: true
|
|
- label: Mining layers use int7 NoisyGEMM
|
|
required: true
|
|
- label: Non-mining layers use int8 vanilla Pearl GEMM
|
|
required: true
|
|
- label: `jarvis mine init --model <pearl-model-id>` succeeds
|
|
required: true
|
|
- label: `jarvis mine start` succeeds
|
|
required: true
|
|
- label: `jarvis ask` succeeds through the mining engine
|
|
required: true
|
|
- label: `jarvis mine status` reports gateway/mining metrics
|
|
required: true
|
|
- label: `jarvis mine validate-model --allow-planned` passes
|
|
required: true
|
|
- label: Gateway/miner logs show no submission errors
|
|
required: true
|
|
- type: textarea
|
|
id: artifacts
|
|
attributes:
|
|
label: Artifacts
|
|
description: Attach logs, metrics, model config, and command output.
|
|
placeholder: |
|
|
- /v1/models output:
|
|
- `jarvis mine status` output:
|
|
- `jarvis mine validate-model --output` JSON:
|
|
- gateway metrics excerpt:
|
|
- miner logs:
|
|
- PR/commit that flips status to validated:
|
|
validations:
|
|
required: true
|