From d74fb68b68630a3d72286f4e5a7494f4fe0e4e5d Mon Sep 17 00:00:00 2001 From: Avanika Narayan Date: Tue, 5 May 2026 20:29:55 -0700 Subject: [PATCH] docs(mining): record Gemma/Qwen Pearl validation blockers Record H100 validation findings for planned Gemma/Qwen Pearl models and keep them planned until public artifacts pass runtime validation. --- docs/development/pearl-model-enablement.md | 19 +++++++++++++++++++ docs/user-guide/mining.md | 6 ++++++ src/openjarvis/mining/_models.py | 21 +++++++++++++++++---- 3 files changed, 42 insertions(+), 4 deletions(-) diff --git a/docs/development/pearl-model-enablement.md b/docs/development/pearl-model-enablement.md index ea6235a4..9b54a200 100644 --- a/docs/development/pearl-model-enablement.md +++ b/docs/development/pearl-model-enablement.md @@ -23,6 +23,23 @@ The current validated model remains: pearl-ai/Llama-3.3-70B-Instruct-pearl ``` +## Current Validation Findings + +The H100 smoke run validated the default Llama Pearl model end to end through +`jarvis mine start`, vLLM `/v1/models`, OpenJarvis inference routing, Pearl +gateway template refresh, and `jarvis mine validate-model`. + +The Qwen and Gemma targets remain planned: + +- `pearl-ai/Qwen3.5-9B-pearl`, `pearl-ai/Qwen3.6-27B-pearl`, and + `pearl-ai/Gemma-4-E4B-it-pearl` are not publicly available artifacts yet. +- `pearl-ai/Gemma-4-31B-it-pearl` exists and selects Pearl mining kernels on + H100, but vLLM fails during Gemma4 multimodal profiling because the published + artifact is missing processor/preprocessor metadata required by Transformers. + A local cache experiment proved the processor can be loaded only after + injecting metadata from `google/gemma-4-31B-it`; that is not sufficient for + OpenJarvis promotion because a clean user install would still fail. + ## Enablement Checklist 1. Reproduce the current Llama Pearl model recipe. @@ -34,6 +51,8 @@ pearl-ai/Llama-3.3-70B-Instruct-pearl 2. Convert the target model. - Start with `Qwen/Qwen3.5-9B`; it is the smallest target. - Generate Pearl-compatible quantized weights and metadata. + - For Gemma4 artifacts, include the base model's processor metadata required + by vLLM's Gemma4 multimodal profiler. - Publish under the planned `pearl-ai/*-pearl` id or a staging namespace. 3. Validate the Pearl vLLM plugin path. diff --git a/docs/user-guide/mining.md b/docs/user-guide/mining.md index 6e02e0f3..f6bbfea0 100644 --- a/docs/user-guide/mining.md +++ b/docs/user-guide/mining.md @@ -88,6 +88,12 @@ pearl-ai/Llama-3.3-70B-Instruct-pearl The Qwen and Gemma targets are tracked in the model registry as planned until Pearl quantization and H100/H200 validation are complete. +Current status: the default Llama Pearl model is the only validated model. +The Gemma 31B Pearl artifact has been seen by OpenJarvis on H100, but it is not +promoted because the published artifact is missing Gemma4 processor metadata +needed for clean vLLM startup. The Qwen and smaller Gemma Pearl artifacts still +need to be published and validated. + When validating a newly converted Pearl model on a mining host, run: ```bash diff --git a/src/openjarvis/mining/_models.py b/src/openjarvis/mining/_models.py index e2ccec41..c7e95a94 100644 --- a/src/openjarvis/mining/_models.py +++ b/src/openjarvis/mining/_models.py @@ -43,7 +43,10 @@ PEARL_MODEL_SPECS: tuple[PearlModelSpec, ...] = ( status="planned", min_vram_gb=24.0, default_max_model_len=8192, - notes="Planned target; validation tracked in open-jarvis/OpenJarvis#316.", + notes=( + "Planned target; Pearl artifact is not publicly available yet. " + "Validation tracked in open-jarvis/OpenJarvis#316." + ), ), PearlModelSpec( model_id="pearl-ai/Qwen3.6-27B-pearl", @@ -51,7 +54,10 @@ PEARL_MODEL_SPECS: tuple[PearlModelSpec, ...] = ( status="planned", min_vram_gb=80.0, default_max_model_len=8192, - notes="Planned target; validation tracked in open-jarvis/OpenJarvis#317.", + notes=( + "Planned target; Pearl artifact is not publicly available yet. " + "Validation tracked in open-jarvis/OpenJarvis#317." + ), ), PearlModelSpec( model_id="pearl-ai/Gemma-4-E4B-it-pearl", @@ -59,7 +65,10 @@ PEARL_MODEL_SPECS: tuple[PearlModelSpec, ...] = ( status="planned", min_vram_gb=24.0, default_max_model_len=8192, - notes="Planned target; validation tracked in open-jarvis/OpenJarvis#318.", + notes=( + "Planned target; Pearl artifact is not publicly available yet. " + "Validation tracked in open-jarvis/OpenJarvis#318." + ), ), PearlModelSpec( model_id="pearl-ai/Gemma-4-31B-it-pearl", @@ -67,7 +76,11 @@ PEARL_MODEL_SPECS: tuple[PearlModelSpec, ...] = ( status="planned", min_vram_gb=80.0, default_max_model_len=8192, - notes="Planned target; validation tracked in open-jarvis/OpenJarvis#319.", + notes=( + "Artifact exists but H100 validation is blocked by missing Gemma4 " + "processor/preprocessor metadata in the published Pearl artifact. " + "Validation tracked in open-jarvis/OpenJarvis#319." + ), ), )