granite-4.2-8b-msvc2008-decomp-lora API & Inference Endpoint

Results (held-out test, 205 functions)

Each function was attempted up to 6 times, with checker feedback between attempts. Sampling used temperature 1.0 and top_p 0.95, with thinking off. A match counts only when the official checker reports a byte-exact match. Test functions and their near-duplicates were excluded from training.

Table with columns: Base Granite 4.2 8B, This adapter
	Base Granite 4.2 8B	This adapter
Exact byte match	7 (3.4%)	26 (12.7%)
Compiled at least once	53 (26%)	131 (64%)
Compiled on first attempt	9 (4%)	62 (30%)
Output tokens used	1.76M	0.62M

The adapter is strong on small functions: 26 of 84 functions under 64 bytes matched. It did not match any test function of 64 bytes or larger.

Training

QLoRA (NF4 base, bf16 compute), rank 64, alpha 64, all attention and MLP projections
1 epoch, learning rate 2e-4 with cosine decay
1,449 generation examples and 1,037 repair examples, about 13.8M tokens
Every target answer was a contributed source that passed a fresh official byte-match check
Loss was computed on the answer only
Prompts are rendered with Granite's chat template and enable_thinking=False
Trained on one RTX 3090 Ti in about 6 hours

Usage (vLLM)

bash

vllm serve ibm-granite/granite-4.2-8b --enable-lora --max-lora-rank 64 \  --lora-modules granite-msvc-decomp=<this-repo>

Send chat requests with chat_template_kwargs: {"enable_thinking": false}. Give the target's assembly, calling-convention hint, compiler flags and the surrounding declarations, and ask for one complete C++ definition in a ```cpp block.

Limitations

Generated code must be checked by recompiling it. A compiling answer is often not an exact match.
It is tuned for one compiler (VC++ 2008 SP1) and one prompt format, so other setups will do worse.

Results (held-out test, 205 functions)

Table with columns: Base Granite 4.2 8B, This adapter
	Base Granite 4.2 8B	This adapter
Exact byte match	7 (3.4%)	26 (12.7%)
Compiled at least once	53 (26%)	131 (64%)
Compiled on first attempt	9 (4%)	62 (30%)
Output tokens used	1.76M	0.62M

The adapter is strong on small functions: 26 of 84 functions under 64 bytes matched. It did not match any test function of 64 bytes or larger.

Training

QLoRA (NF4 base, bf16 compute), rank 64, alpha 64, all attention and MLP projections
1 epoch, learning rate 2e-4 with cosine decay
1,449 generation examples and 1,037 repair examples, about 13.8M tokens
Every target answer was a contributed source that passed a fresh official byte-match check
Loss was computed on the answer only
Prompts are rendered with Granite's chat template and enable_thinking=False
Trained on one RTX 3090 Ti in about 6 hours

Usage (vLLM)

bash

vllm serve ibm-granite/granite-4.2-8b --enable-lora --max-lora-rank 64 \  --lora-modules granite-msvc-decomp=<this-repo>

Limitations

Generated code must be checked by recompiling it. A compiling answer is often not an exact match.
It is tuned for one compiler (VC++ 2008 SP1) and one prompt format, so other setups will do worse.

granite-4.2-8b-msvc2008-decomp-lora

README

Results (held-out test, 205 functions)

Training

Usage (vLLM)

Limitations

Explore FriendliAI today

README

Results (held-out test, 205 functions)

Training

Usage (vLLM)

Limitations