Yanbo Liang

str(weakref) throws error `KeyError: '__name__'` if the original obj override __getattr__

6

# Bug report A clear and concise description of what the bug is. Include a minimal, reproducible example (https://stackoverflow.com/help/minimal-reproducible-example), if possible. Repo: ``` class MyConfig(dict): def __getattr__(self, x): return self[x]...

[Dynamo] Support context wrapping(e.g, torch.no_grad) on nested functions w/o closure

1

Fixes 14k github models: https://github.com/jansel/pytorch-jit-paritybench/blob/master/generated/test_ELEKTRONN_elektronn3.py cc @mlazos @soumith @voznesenskym @penguinwu @anijain2305 @EikanWang @jgong5 @Guobing-Chen @chunyuan-w @XiaobingSuper @zhuhaozhe @blzheng @Xia-Weiwen @wenzhe-nrv @jiayisunx @desertfire

[Dynamo] Supports if condition on user defined object

3

Fixes Meta internal user case, see the pattern in unit test. cc @mlazos @soumith @voznesenskym @penguinwu @anijain2305 @EikanWang @jgong5 @Guobing-Chen @chunyuan-w @XiaobingSuper @zhuhaozhe @blzheng @Xia-Weiwen @wenzhe-nrv @jiayisunx @desertfire

This is based on #57. Please checkout https://github.com/yanboliang/gpt-fast/tree/mixtral-moe to try this. Performance numbers (tokens/second): ``` | | 1 GPU | 2 GPU | 8 GPU | |------------------|---------|-----------|-------------| |baseline(bfloat16)| OOM |...

[Dynamo] Fix guards for script_if_tracing or lru_cache fn with default args

2

Fixes #120387 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @aakhundov @kadeng

Add templated attention BLOCK_M & BLOCK_N default size for different head_dim

1

Run different head_dims [64, 128], which are the most popular ones across major GPT models. Enumerate different ```BLOCK_M``` and ```BLOCK_N``` candidates [16, 32, 64, 128], and get the best config...

Dynamo doesn't log backward graphs in compilation metrics

3

### 🐛 Describe the bug Repro: ``` import torch import torch.nn.functional as F class BasicModule(torch.nn.Module): def __init__(self): super().__init__() self.linear1 = torch.nn.Linear(10, 10) self.scale = torch.randn(1, 10) def forward(self, x): x...

[Inductor] default block size for head_dim = 256 for flex attention

1

## H100 ### torch.bfloat16 No major change, as expected. ``` | Type | Speedup | batch_size | num_heads | q_seq_len | k_seq_len | head_dim | score_mod | dtype | |---------|-----------|--------------|-------------|-------------|-------------|------------|-------------|----------------|...

[Not for landing][Prototype] Add torch._record_scalar

4

I'm adding this prototype to demonstrate a possible way that how intermediate logging works with torch.compile. The following example is a minimized repro and how it ladders up with proposed...

Add Inductor micro benchmark workflow

2

Fixes #ISSUE_NUMBER