GitLab

Co-authored-by: moshuosha <moshuosha@qq.com>

Update handling for KeyError in state_dict.pop() for non-existing keys.
Changed state_dict.pop(f"h.{d}.attn.bias") to state_dict.pop(f"h.{d}.attn.bias", None) to prevent KeyError exceptions.


The following code can re-produce the issue
```
from transformers import AutoTokenizer, GPT2Model, GPT2Config
from flash_attn.models.gpt import GPTLMHeadModel, GPTModel

# >>> transformers.__version__
# '4.38.2'

model_path = 'gpt2'
output_model_path = 'gpt2_model'
config = GPT2Config.from_pretrained(model_path, output_hidden_states=True)
model = GPT2Model.from_pretrained(model_path, from_tf=False, config=config)
'''
model fine-tuning here
'''
# dump the fine-tuned model
model.save_pretrained(output_model_path)

# load the fine-tuned model
config = GPT2Config.from_pretrained(output_model_path, output_hidden_states=True)
model = GPTModel.from_pretrained(output_model_path, config=config, strict=True)  # failed due to KeyError: 'h.0.attn.bias'
model = GPTLMHeadModel.from_pretrained(output_model_path, config=config, strict=True)  # failed due to KeyError: 'h.0.attn.bias'

```

When user send `out` as a parameter of the function
`seqlenq_ngroups_swapped` with parameters that trigger,
the CHECK_SHAPE is incorrect (since q shape is modified.)

Co-authored-by: zl <zl@deepseek.com>

* Support unpadded LSE layout.
Co-authored-by: Xinfeng Xie <xfxie.ceca@gmail.com>
Co-authored-by: Jianyu Huang <hjyahead@gmail.com>

* Cleanup

* Fix unpadded LSE on split-kv path

* Fix formatting and comments

* Fix inline vs forceinline

---------
Co-authored-by: Xinfeng Xie <xfxie.ceca@gmail.com>
Co-authored-by: Jianyu Huang <hjyahead@gmail.com>

When timeout connecting, you get URLError: <urlopen error timed out>, In that case, build it from source.

Set `packaging` and `ninja` as build time dependencies rather than runtime dependencies.

All integer parameters are specialized by default, so the two parameters
removed in this commit could lead to kernel re-compilation, even if
they were completely unused.

* Enable paged attention in varlen forward

* Format + fix padding

Optimize compile to 1: avoid oom 2: minimize swap usage 3: avoid threads starvation when letting ninja decide how many workers to spawn or manual MAX_JOBS "guesses". Logic is to take the min value of MAX_JOBS auto-calculated by two metrics: 1: cpu cores 2: free memory. This should allow flash-attn to compile close to the most efficient manner under any consumer/server env. (#832)