Qwen3.8-27B-pi — SVD-LoRA (base: Qwen3.8-27B)

The weight difference between bytkim/Qwen3.8-27B-pi and stock Qwen/Qwen3.8-27B, compressed by exact truncated SVD into a rank-64 LoRA. Load it on any Qwen3.8-27B GGUF with --lora to get the Pi coding-agent behaviour without shipping a second 55 GB model.

What is Qwen3.8-27B-pi?

A Qwen3.8-27B fine-tune built for the Pi agent harness: the iterative loop of reading a repository, editing files, running tools, and responding to their feedback. From the source card:

Supervised fine-tuning on filtered, successful Pi sessions — complete coding workflows, not isolated snippets.
GRPO (reinforcement learning) on top, emphasising completed, checked implementations over generated text.
MTP (multi-token prediction) head, vision support, adjustable reasoning effort.

How this LoRA was made

Download both bytkim/Qwen3.8-27B-pi and Qwen/Qwen3.8-27B (both 1,199 tensors, identical qwen3_5 architecture — no key remapping needed).
Compute Δ = W_pi − W_base per tensor; 496 of 1,199 tensors have a non-zero delta (≥ 0.001 norm). lm_head and embed_tokens are untouched by the fine-tune and are excluded.
Run an exact truncated SVD of every delta at rank 64 (rank 48 where the matrix is only 48×5120 — the GDN in_proj_a/in_proj_b gates, which reconstruct exactly).
Store lora_A = Σ_r^½ V_rᵀ, lora_B = U_r Σ_r^½ in PEFT format.

Nothing is trained; no dataset is used. The adapter is a pure compression of the fine-tune's weight delta.

Measured fidelity

property	value
tensor pairs	496
rank	64 (48 for GDN gate pairs)
alpha	64 (scale 1.0 at inference)
energy retained (mean)	0.549
energy retained (min / max)	0.286 / 1.000
relative reconstruction error (mean)	0.598
target modules	`q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj, in_proj_qkv, in_proj_z, in_proj_a, in_proj_b, out_proj`
dtype	float32

The fine-tune's delta is broad and low-amplitude (max Δ-norm 0.139), so rank 64 captures roughly half the delta energy on average. Expect the Pi behaviour — terser, tool-driven coding — not a bit-exact copy of the source model.

Files

File	Format	Notes
`adapter_model.safetensors`	PEFT LoRA	496 pairs, r=64, alpha=64, fp32
`adapter_config.json`	JSON	`base_model: Qwen/Qwen3.8-27B`
`pi_qwen3.8-27B_svd_lora.gguf`	llama.cpp LoRA GGUF	800 tensors; GDN `in_proj_a`/`in_proj_b` pairs omitted (see note)

Note on the GGUF file: convert_lora_to_gguf.py does not map the GDN in_proj_a/in_proj_b tensors, so the GGUF carries 800 of the 992 adapter tensors. The PEFT adapter is complete (all 496 pairs). If you need the GDN gates, use the PEFT adapter in Python.

Usage

llama.cpp

llama-cli -m Qwen3.8-27B.gguf \
    --lora pi_qwen3.8-27B_svd_lora.gguf \
    -p "Refactor this function and run the tests."

Python (PEFT)

from peft import PeftModel
from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.8-27B",
                                             torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(model, "hotdogs/qwen3.8-27B-pi_svd_lora")

License

Apache-2.0 (upstream Qwen3.8-27B-pi). Base model Qwen3.8-27B is Apache-2.0. Credit: bytkim for Qwen3.8-27B-pi.