Qwen3.5

pllm.qwen3_5_text.v1 lowers the text decoder in Qwen/Qwen3.5-4B revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a. The upstream configuration SHA-256 is ddc63e1c717afa86c865bb5e01313d89d72bb53b97ad4a8a03ba8510c0621670; the checked fixture adds one repository newline and is verified against both representations.

The plan retains the exact 3:1 linear/full layer schedule. Linear layers represent fused input projection, causal convolution state, Q/K L2 normalization, gated-delta decay and update, recurrent matrix state, and gated RMS normalization. Full-attention layers represent doubled query projection, Q/K normalization, partial multimodal rotary sections, KV state, and output gating. MTP configuration remains identity-bound but outside the base decoder graph.

These newly explicit operators currently report missing compiler coverage. Semantic lowering does not claim an executable Qwen3.5 runtime, checkpoint importer, generation quality, or protected recurrent protocol.

Python SDK example

from pathlib import Path

from pllm import lower_model

fixture = Path("crates/pllm-models/tests/fixtures/Qwen3.5-4B-851bf6e-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.0.linear_attention.gated_delta_rule"]["operator"] == "gated_delta_rule"
assert operations["layer.3.self_attention.q_rope"]["attributes"]["mrope_section"] == (11, 11, 10)

API: Python SDK objects and signatures

Run this example from a PLLM source checkout; package installations do not include the pinned fixture. Inspecting its plan is not an execution claim.