Tensormorph/distillgpt2

distilgpt2: export fixtures

35 variants: 34 ok, 0 mismatch, 0 failed, 1 skipped. Generated by python -m tools.export_formats.run_all; machine-readable twin: manifest.json (sha256 per file).

Every file was loaded back with its own runtime (PyTorch, onnxruntime, gguf-py, Keras/TF, TFLite, Core ML, OpenVINO) and its logits on the fixed [2, 8] ids compared with the PyTorch model. max rel diff is relative to max |logit|; ok means within the variant's tolerance (1e-4 for float32, looser for fp16 / int8 / 4-bit). graph says what the file holds besides weights.

format	variant	files	size	graph	status	max rel diff	notes
pytorch	state_dict (our names, zip) .pt	pytorch/distilgpt2.state_dict.pt	312.5 MiB	none	ok	0.0e+00
pytorch	state_dict (HF names, zip) .pth	pytorch/distilgpt2.hf_state_dict.pth	336.5 MiB	none	ok	0.0e+00
pytorch	training checkpoint .pt	pytorch/distilgpt2.training_checkpoint.pt	937.5 MiB	none	ok	0.0e+00	{model: state_dict with 'module.' prefix, optimizer: AdamW state, epoch, global_step}
pytorch	lightning-style .ckpt	pytorch/distilgpt2.lightning_style.ckpt	937.5 MiB	none	ok	0.0e+00	PyTorch zip pickle with a .ckpt extension; layout mimics PyTorch Lightning (Lightning itself not used)
pytorch	full model pickle .pt	pytorch/distilgpt2.full_model_pickle.pt	318.5 MiB	pickled classes	ok	0.0e+00	torch.save(model): needs the Python classes and arbitrary unpickling; an importer must refuse it
torchscript	torch.jit.trace .pt	pytorch/distilgpt2.torchscript.pt	318.6 MiB	TorchScript code	ok	0.0e+00
torch.export	ExportedProgram .pt2	pytorch/distilgpt2.pt2	319.0 MiB	ATen graph	ok	0.0e+00
binary	pytorch_model.bin (zip pickle)	pytorch/pytorch_model.bin	336.5 MiB	none	ok	0.0e+00
binary	pytorch_model.bin (legacy pickle)	pytorch/pytorch_model.legacy.bin	336.5 MiB	none	ok	0.0e+00	pre-1.6 non-zip pickle (copy of Hugging Face's original file, storages on cuda:0)
safetensors	HF names, lm_head omitted (tied)	pytorch/model.safetensors	336.5 MiB	none	ok	0.0e+00
onnx	dynamo exporter, static [2,8]	onnx/distilgpt2.onnx	318.9 MiB	ONNX graph	ok	2.3e-06	weights inline in the .onnx
onnx	dynamo exporter, dynamic axes	onnx/distilgpt2.dynamic.onnx onnx/distilgpt2.dynamic.onnx.data	318.9 MiB	ONNX graph	ok	2.6e-06	symbolic batch / sequence; weights in .onnx.data (exporter default)
onnx	legacy exporter, opset 17	onnx/distilgpt2.legacy_opset17.onnx	460.9 MiB	ONNX graph	ok	2.6e-06	TorchScript-based exporter (dynamo=False): Gemm/MatMul patterns differ
onnx	external data (.onnx + .bin)	onnx/distilgpt2.external.onnx onnx/distilgpt2.external.bin	637.4 MiB	ONNX graph	ok	2.3e-06	all initializers in one sibling .bin file
onnx	dynamic int8 (MatMulInteger)	onnx/distilgpt2.int8_dynamic.onnx	116.8 MiB	ONNX graph	ok	1.1e-01	onnxruntime quantize_dynamic on the opset-17 model: per-channel int8 weights; activations quantized per tensor at run time, so lossy by design (DistilGPT2: ~11% of max
gguf	f32	gguf/distilgpt2.f32.gguf	312.5 MiB	none (arch + hyperparameters)	ok	0.0e+00
gguf	f16 (1-D tensors f32)	gguf/distilgpt2.f16.gguf	156.4 MiB	none (arch + hyperparameters)	ok	3.1e-04
gguf	q8_0 (1-D tensors f32)	gguf/distilgpt2.q8_0.gguf	83.2 MiB	none (arch + hyperparameters)	ok	1.1e-02	block-quantized: 32 values per 34-byte block
gguf	q4_0 (1-D tensors f32)	gguf/distilgpt2.q4_0.gguf	44.1 MiB	none (arch + hyperparameters)	ok	2.0e-01	block-quantized: 32 values per 18-byte block
keras	rebuilt model (in memory)	—	—	-	ok	2.2e-06	reference for the files below; nothing written
keras	Keras 3 .keras (zip)	tensorflow/distilgpt2.keras	312.7 MiB	Keras config	ok	2.2e-06	zip: config.json + metadata.json + model.weights.h5
hdf5	Keras full model .h5	tensorflow/distilgpt2.h5	312.7 MiB	Keras config (model_config attr)	ok	2.2e-06	legacy HDF5 format
hdf5	Keras weights only .weights.h5	tensorflow/distilgpt2.weights.h5	312.7 MiB	none	ok	2.2e-06	no architecture: needs the code or a template
tf-checkpoint	TF2 object checkpoint .ckpt	tensorflow/distilgpt2.ckpt.index tensorflow/distilgpt2.ckpt.data-00000-of-00001	312.5 MiB	none (object graph only)	ok	2.2e-06
tf-checkpoint	TF1 checkpoint .ckpt + .meta	tensorflow/distilgpt2_tf1.ckpt.index tensorflow/distilgpt2_tf1.ckpt.data-00000-of-00001 tensorflow/distilgpt2_tf1.ckpt.meta tensorflow/checkpoint	312.7 MiB	TF1 GraphDef in .meta	ok	2.3e-06	variable names like OpenAI's GPT-2 release (model/h0/attn/c_attn/w); weights are [in, out]
savedmodel	SavedModel directory	tensorflow/distilgpt2_saved_model	625.6 MiB	TF GraphDef (saved_model.pb)	ok	2.2e-06
tflite	float32	tensorflow/distilgpt2.float32.tflite	309.6 MiB	TFLite subgraph	ok	7.5e-07	static input [2, 8]
tflite	float16 weights	tensorflow/distilgpt2.float16.tflite	154.8 MiB	TFLite subgraph	ok	4.1e-04	static input [2, 8]; weights fp16 + DEQUANTIZE ops
tflite	dynamic-range int8 weights	tensorflow/distilgpt2.int8_dynamic.tflite	115.5 MiB	TFLite subgraph	ok	5.4e-02	static input [2, 8]; int8 weights, float activations
coreml	mlprogram FP32 .mlpackage	coreml/distilgpt2.mlpackage	309.8 MiB	MIL program	ok	1.2e-06	directory: Manifest.json + Data/com.apple.CoreML/{model.mlmodel, weights/weight.bin}
coreml	mlprogram FP16 .mlpackage	coreml/distilgpt2.fp16.mlpackage	154.9 MiB	MIL program	ok	1.1e-03	weights and compute in fp16
coreml	neuralnetwork .mlmodel	coreml/distilgpt2.mlmodel	456.9 MiB	NeuralNetwork layers	ok	1.0e-06	single protobuf file, weights inline
openvino	IR fp32 (.xml + .bin)	openvino/distilgpt2.xml openvino/distilgpt2.bin	313.7 MiB	OpenVINO IR graph (.xml)	ok	1.2e-06
openvino	IR fp16-compressed weights	openvino/distilgpt2.fp16.xml openvino/distilgpt2.fp16.bin	157.5 MiB	OpenVINO IR graph (.xml)	ok	4.1e-04	weights stored fp16 in the .bin, decompressed by Convert ops in the graph
tensorrt	.engine from ONNX	—	—	—	skipped	—	needs TensorRT + an NVIDIA CUDA GPU (tensorrt importable: False, CUDA: False); on such a machine run this script or: trtexec --onnx=exports/distilgpt2/onnx/distilgpt2.onnx --saveEngine=exports/distilgpt2/tensorrt/distilgpt2.engine --exportLayerInfo=exports/distilgpt2/tensorrt/distilgpt2.layers.json --profilingVerbosity=detailed

Environment: torch 2.14.1, onnx 1.23.2, onnxscript 0.7.2, onnxruntime 1.31.0, gguf 0.19.0, tensorflow 2.21.0, keras 3.15.1, coremltools 9.0, openvino 2026.4.1, safetensors 0.8.0, numpy 2.5.4; macOS-26.6.2-arm64-arm-64bit, Python 3.12.11.