pllmModels
Gemma 4
Variant-parameterized text-decoder lowering for exact Gemma 4 E2B and E4B configurations.
PLLM documentationView Markdown ↗
The Gemma 4 adapter validates exact outer text-decoder contracts for:
| Adapter | Repository revision | Raw configuration SHA-256 |
|---|---|---|
pllm.gemma4_e2b_text.v1 | google/gemma-4-E2B-it@3e22461f65e89153144f8adb70e3b8c2cc9845a7 | 1b28f3d2c3100f6c594754b81107428bd7b822a7f48272ca681dae9d2ec38330 |
pllm.gemma4_e4b_text.v1 | google/gemma-4-E4B-it@ee0ef6023621cff504d758262d4e04895a5af4a2 | 33b10c02df3c2e8536cf323d29d53262aaa2f4d11dbe19bc729373fbe90295d4 |
Both plans preserve alternating local/global attention, two head widths, Q/K normalization, PLE, shared-KV suffix state, GeLU-tanh branches, branch-output normalization, and final softcap. E2B additionally doubles effective MLP width in its shared suffix. Variant dimensions and layer schedules are data-driven; compiler paths remain model-neutral.
Python SDK example
from pathlib import Path
from pllm import lower_model
fixture = Path("crates/pllm-models/tests/fixtures/gemma-4-E2B-it-3e22461f-config.json")
plan = lower_model(fixture.read_bytes(), batch=1, max_input_tokens=16, max_new_tokens=4)
operations = {operation["id"]: operation for operation in plan.prefill["operations"]}
assert operations["layer.14.gate_proj"]["output_shape"][-1] == 6144
assert operations["layer.15.gate_proj"]["output_shape"][-1] == 12288
assert operations["layer.34.attention_scores"]["inputs"][1] == "layer.14.key_append"API: Python SDK objects and signatures
Run this example from a PLLM source checkout; package installations do not include the pinned E2B fixture. Lowering does not execute weights.