// Knowledge distillation techniques for model compression: logit-level, feature-level, and relation-based distillation. KD-Lib library and practical workflows for training student models.
Prepare and validate Parameter Golf record folders: self-contained train_gpt.py, README.md, submission.json, FineWeb SP1024 BPB accounting, artifact-size logging, run logs, and PR-ready folder hygiene.
Run Parameter Golf competition submissions on RunPod GPU Pods. Covers required operator inputs, RunPod pod specs, FineWeb SP1024 data caching, record-folder hygiene, torchrun launch commands, monitoring, artifact-size checks, and result collection.
Hands-on implementation template and API reference for writing, tuning, debugging, and benchmarking Triton GPU kernels. Covers the full triton.language API surface, autotuning patterns, profiling workflows, and production integration.
Hands-on implementation template and API reference for writing, tuning, debugging, and benchmarking Triton GPU kernels. Covers the full triton.language API surface, autotuning patterns, profiling workflows, and production integration.
Tencent AngelSlim — accessible, comprehensive, and efficient toolkit for large model compression. Quantization (FP8/INT4/NVFP4/1.25-bit), pruning, speculative decoding (Eagle3), and diffusion model compression.
DistilQwen2.5 — Alibaba's industrial practices for training distilled open lightweight language models. Knowledge distillation from Qwen2.5 72B into smaller 0.5B-7B models.
| name | knowledge-distillation |
| description | Knowledge distillation techniques for model compression: logit-level, feature-level, and relation-based distillation. KD-Lib library and practical workflows for training student models. |
| tags | ["knowledge-distillation","model-compression","student-teacher","kd-lib","pytorch","zorai"] |
Knowledge distillation transfers knowledge from a larger teacher model to a smaller student model. Combined with pruning and quantization, it forms the critical middle step in the P-KD-Q compression pipeline.
uv pip install kd-lib
from kd_lib import distill
import torch.nn.functional as F
def kd_loss(student_logits, teacher_logits, labels, temperature=4.0, alpha=0.5):
soft_targets = F.softmax(teacher_logits / temperature, dim=-1)
soft_prob = F.log_softmax(student_logits / temperature, dim=-1)
kd = F.kl_div(soft_prob, soft_targets, reduction="batchmean") * (temperature ** 2)
ce = F.cross_entropy(student_logits, labels)
return alpha * kd + (1 - alpha) * ce
| Method | What it transfers | Best For |
|---|---|---|
| Logit distillation | Output probability distribution | Classification, generation |
| Feature distillation | Intermediate hidden states | Transformer layers |
| Relation distillation | Relationships between representations | Structured outputs |
| Self-distillation | Model teaches itself | No teacher needed |
| Online distillation | Teacher & student train jointly | Both models improve |