GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Shreyansh327
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Shreyansh327
Model Tree
Qwen/Qwen2.5-Coder-7B-Instruct
Input Modalities
Output Modalities
Supported Functionality
Evaluated on 120 episodes per arm (same tasks, same seeds, 4 samples per task), H100, sampled with the same settings as training.
| Split | Base | SFT | SFT + RL (this model) |
|---|---|---|---|
| All 120 episodes, correct | 19/120 | 18/120 | 24/120 |
| Trained tasks (16 ep) | 81.2% | 68.8% | 93.8% |
| Untrained tasks, seen split (32 ep) | 15.6% | 21.9% | 21.9% |
| Held-out level 1 | 0% | 0% | 0% |
| Held-out level 2 (fusion) | 2.8% | 0% | 5.6% |
Paired base → RL deltas with task-level bootstrap 95% CIs (percentage points):
| Split | Δ | 95% CI |
|---|---|---|
| Trained | +12.5 | [0, 37.5] |
| Untrained | +6.2 | [−12.5, 25.0] |
| Held-out L1 | 0 | — |
| Held-out L2 | +2.8 | [0, 8.3] |
Read this honestly: RL helped most on the tasks it trained on, the gain on held-out tasks is small, and every CI includes zero. This is a small-sample research run, not a strong kernel-writing model. Speedups over PyTorch were rare; most correct kernels run at roughly reference speed.
Qwen/Qwen2.5-Coder-7B-Instruct @ c03e6d358207e414f1eca0bb1891e29f1db0e242423217d9) on multiple random inputs;
output and state tampering detected with keyed GPU fingerprintspython
from transformers import AutoModelForCausalLM, AutoTokenizer repo = "Shreyansh327/Qwen2.5-Coder-7B-KernelEnv-Fusion-RL"tok = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="bfloat16", device_map="auto")Prompt it with a KernelBench reference module and ask for a ModelNew class that uses Triton kernels.
The RL adapter was trained on the SFT-merged model, so merge in order:
python
import torchfrom transformers import AutoModelForCausalLMfrom peft import PeftModel repo = "Shreyansh327/Qwen2.5-Coder-7B-KernelEnv-Fusion-RL"m = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-Coder-7B-Instruct", revision="c03e6d358207e414f1eca0bb1891e29f1db0e242", torch_dtype=torch.bfloat16)m = PeftModel.from_pretrained(m, repo, subfolder="adapters/sft").merge_and_unload()m = PeftModel.from_pretrained(m, repo, subfolder="adapters/rl").merge_and_unload()| File | sha256 |
|---|---|
adapters/sft/adapter_model.safetensors | 478d5412a776c35674c150b7525801d25dd947b8a76bb4bf7162a5c66967c2b4 |
adapters/rl/adapter_model.safetensors | 86df23e5a587f2d8ab25bdea2387bdfd20d3ffd70b244589051c7a8934575b38 |
Full per-arm metrics: eval/comparison.json.
from transformers import AutoModelForCausalLM, AutoTokenizer repo = "Shreyansh327/Qwen2.5-Coder-7B-KernelEnv-Fusion-RL"tok = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="bfloat16", device_map="auto")import torchfrom transformers import AutoModelForCausalLMfrom peft import PeftModel repo = "Shreyansh327/Qwen2.5-Coder-7B-KernelEnv-Fusion-RL"m = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-Coder-7B-Instruct", revision="c03e6d358207e414f1eca0bb1891e29f1db0e242", torch_dtype=torch.bfloat16)m = PeftModel.from_pretrained(m, repo, subfolder="adapters/sft").merge_and_unload()m = PeftModel.from_pretrained(m, repo, subfolder="adapters/rl").merge_and_unload()