qwen3-14b-code-agent-v2 — fully merged code-agent model

v2 = qwen3-14b-code-agent (fine-tune) + wsurgery weight patch + MCP LoRA adapter, all merged into plain safetensors. Load it like any normal HF model — no patch step, no adapter directory:

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
    "harshagowtham/qwen3-14b-code-agent-v2", torch_dtype="bfloat16", device_map="auto")

or with the engine (auto 4-bit on small GPUs, OpenAI-compatible API, thinking mode, tool calls):

from llm_engine import LocalLLM   # https://huggingface.co/harshagowtham/qwen-agent-scripts
llm = LocalLLM("harshagowtham/qwen3-14b-code-agent-v2", hf_token="hf_...")

Recipe

step	source	operation
1	harshagowtham/qwen3-14b-code-agent	starting weights (Qwen3-14B fine-tune)
2	harshagowtham/qwen3-14b-code-agent-wsurgery	W = W_base + 0.70*(W_ft - W_base) on the top-8 o_proj + top-5 down_proj residual writers (13 tensors, byte-exact splice)
3	harshagowtham/qwen3-14b-code-agent-mcp-lora	LoRA merge_and_unload() (tool-calling adapter)

Why: the fine-tune drifted lightly but uniformly across all 280 projection tensors; the surgery trims 30% of that drift on the 13 highest-energy residual writers (deep layers dominate), keeping the learned code/tool behaviour. The MCP adapter restores/boosts structured tool calling. Full analysis: harshagowtham/qwen3-14b-code-agent-wsurgery -> SURGERY_REPORT.md.

Build it yourself (Kaggle, ~40-60 min)

Two equivalent paths:

1. Notebook (recommended): open kaggle_notebooks/build_v2_merged_weights.ipynb in a Kaggle notebook with GPU T4 x2, Internet ON, and your HF token.
2. CLI script:

pip install -U transformers accelerate peft huggingface_hub sentencepiece
python build_v2.py --token hf_YOUR_WRITE_TOKEN
# options: --workdir /kaggle/working/v2-build --skip-push --device auto

Requirements: ~28-29 GB RAM (bf16 CPU merge — Kaggle has ~29 GB), ~35 GB free disk. The script verifies every step (patch sha256s, adapter merge, 50-token smoke test) before pushing to harshagowtham/qwen3-14b-code-agent-v2.

Disk: how the build avoids "No space left on device"

A naive build needs 28 GB (download) + 28 GB (save) = 56 GB at the same time, which ENOSPCs on a ~58 GB Kaggle volume with anything else on it — typically dying at Writing model shards: 0%. This build instead:

1. reuses an already-cached snapshot when re-running in the same session (no 28 GB re-download after a crashed attempt);
2. frees the 28 GB fine-tune snapshot right after the LoRA merge — hub-loaded safetensors are mmap-backed and on Linux an unlinked mapping stays readable, so save_pretrained streams through the still-live mapping. Peak disk ~30 GB instead of ~56 GB;
3. checks free space before the save and prints free/needed clearly;
4. removes stale partial shards from a crashed run before saving;
5. after a successful push, deletes the merged dir and hub caches (Kaggle caps persisted session output at 20 GB); --keep-output keeps them.

Troubleshooting: ImportError: Found an incompatible version of torchao

Some Kaggle images ship torchao 0.10 while the installed PEFT requires

= 0.16 whenever torchao is present, which crashes PeftModel.from_pretrained (step 4/6). As of this commit build_v2.py auto-applies a guard that tells PEFT "torchao unavailable", so PEFT skips its torchao LoRA dispatch and uses the standard LoRA path. If you hit the error with an old copy: git pull / re-download build_v2.py and re-run — no pip changes needed (do NOT pip uninstall torchao; other packages on the image may want it).

Files

build_v2.py                            # standalone CLI builder (this repo)
kaggle_notebooks/
  build_v2_merged_weights.ipynb        # same build as a Kaggle notebook

Model card

• Base: Qwen/Qwen3-14B (40 layers, hidden 5120, 32k native context)
• Languages: English (code + general)
• Intended use: coding agent, tool calling, long-form site/code generation
• Serving: transformers / llm_engine / vLLM (standard Qwen3 arch — enable_thinking chat-template flag supported)
• Quantization-friendly: works with bitsandbytes NF4 (auto in llm_engine under 24 GB VRAM)

License

Apache-2.0, inherited from Qwen3-14B and the constituent repos.