Method x Model Compatibility¶

Generated from tl.report.doctor on representative eager PyTorch models. These rows are a smoke reference for ordinary dense eager execution, not a complete certification matrix. Environment rows are excluded from this projection; the full doctor report includes them.

Generation snippet:

import torch
from torch import nn
import torchlens as tl

models = {
    "linear_mlp": nn.Sequential(nn.Linear(4, 6), nn.ReLU(), nn.Linear(6, 2)).eval(),
    "conv_pool": nn.Sequential(
        nn.Conv2d(1, 2, 3), nn.ReLU(), nn.AdaptiveAvgPool2d((1, 1)), nn.Flatten()
    ).eval(),
}
inputs = {
    "linear_mlp": torch.ones(1, 4),
    "conv_pool": torch.ones(1, 1, 5, 5),
}

for name, model in models.items():
    print(name)
    print(tl.report.doctor(model, inputs[name]).to_markdown())

Representative Results¶

Both representative models report pass for every check:

• HF Transformers wrapper
• Accelerate device_map='auto'
• Accelerate CPU/disk offload
• bitsandbytes 8-bit/4-bit
• GPTQ/AWQ quantization
• CUDA graphs
• Serving engines (vLLM/SGLang)
• PEFT/LoRA adapters
• Tied/shared parameters
• Multi-GPU RNG
• nn.DataParallel
• DistributedDataParallel
• FSDP
• DTensor / sharded tensors
• Device mesh
• Tensor parallel (TP)
• Pipeline parallel (PP)
• DeepSpeed
• torch.compile
• FX GraphModule
• Protocol-invisible belt coverage
• Lightning training_step mid-loop
• vmap/functorch
• Quantized tensors/modules
• fp8 (float8_*) tensors
• DeviceContext factory injection
• Single-thread design

Interpretation: plain eager dense PyTorch models are the compatibility baseline. For wrappers, compiled execution, sharding/offload, quantization, or concurrent capture, run tl.report.doctor(model, x) on the exact model/input pair and include the report when filing an issue.