Skip to content

reland - use fast softmax only on prefill#309

Closed
jaygala223 wants to merge 10 commits into
HabanaAI:v1.17-synapsefrom
jaygala223:jay-softmax-fix-2
Closed

reland - use fast softmax only on prefill#309
jaygala223 wants to merge 10 commits into
HabanaAI:v1.17-synapsefrom
jaygala223:jay-softmax-fix-2

Conversation

@jaygala223
Copy link
Copy Markdown

What does this PR do?

Fixes # (issue)

Before submitting

  • This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
  • Did you make sure to update the documentation with your changes?
  • Did you write any new necessary tests?

pramodkumar-habanalabs and others added 10 commits July 11, 2024 11:44
- Update attention module and usages
- Add --flash_attention_fp8 flag
…" (#295)

This reverts commit 35f6fbe.

Co-authored-by: Eran Geva <egeva@habana.ai>
* TP reference -  ibm foundation-model-stack

* Code cleanup -removed unused code

---------

Co-authored-by: Kalyan <kkumar@habana.ai>
* Support loading checkpoint with INC

* load_cp explanation

* Add torch_dtype bf16 for the model

---------

Co-authored-by: yan tomsinsky <ytomsinsky@habana.ai>
@jaygala223 jaygala223 closed this Jul 18, 2024
gplutop7 pushed a commit that referenced this pull request Oct 15, 2025
* Merge v1.18-release

* Hot fix regional compilation (huggingface#2005)

Co-authored-by: regisss <15324346+regisss@users.noreply.github.com>

* Enable mixtral 8x7b accuracy evaluation (huggingface#1986)

Co-authored-by: Rafal <rbogdanowicz@habana.ai>

* Update readme files for explicit lazy mode (huggingface#1921)

Co-authored-by: Karol Brejna <karol.brejna@intel.com>
Co-authored-by: Piotr Bielak <piotr.bielak@intel.com>

* [llama-vision] Remove token_idx_cpu parameter (huggingface#2018)

Integer parameter token_idx_cpu passed to mllama's forward()
method caused an issue with hpu graph cache which led to
performance drop.

Signed-off-by: Urszula <urszula.golowicz@intel.com>

* Update README examples (huggingface#2020)

* Fix examples in README

audio-classification:
- add space between "False" and backslash

image-to-text:
- add "datasets" to requirements.txt

pytorch-image-models:
- add "datasets" to requirements.txt

sentence-transformers-training/nli:
- add command to properly discover HABANA_VISIBLE_MODULES

sentence-transformers-training/sts:
- add command to properly discover HABANA_VISIBLE_MODULES

speech-recognition:
- add `--trust_remote_code` for seq2seq examples

stable-diffusion/training:
- add missing OpenCV requirement for ControlNet Training

Co-authored-by: Karol Brejna <karol.brejna@intel.com>

* Review fixes: remove grabbing all modules

---------

Co-authored-by: Karol Brejna <karol.brejna@intel.com>
Co-authored-by: karol-brejna-i <karolbrejna@apache.org>

* Pin latest optimum to force mutual updates (huggingface#2016)

pin latest optimum to force mutual updates

* Fix FP8 support and address related issues (huggingface#2010)

- Resolve bugs related to FP8 (floating point 8-bit) computation
- Improve stability and correctness of FP8 operations
- Add/fix tests to validate FP8 functionality
- Update relevant documentation and comments

Co-authored-by: IlyasMoutawwakil

---------

Signed-off-by: Urszula <urszula.golowicz@intel.com>
Co-authored-by: Adam Stachowicz <astachowicz@habana.ai>
Co-authored-by: Ilyas Moutawwakil <57442720+IlyasMoutawwakil@users.noreply.github.com>
Co-authored-by: regisss <15324346+regisss@users.noreply.github.com>
Co-authored-by: Rafal Bogdanowicz <rafal.bogdanowicz@intel.com>
Co-authored-by: Rafal <rbogdanowicz@habana.ai>
Co-authored-by: Jan Kamiński <jkaminski@habana.ai>
Co-authored-by: Karol Brejna <karol.brejna@intel.com>
Co-authored-by: Piotr Bielak <piotr.bielak@intel.com>
Co-authored-by: Urszula Golowicz <urszula.golowicz@intel.com>
Co-authored-by: Piotr Bielak <pbielak@users.noreply.github.com>
Co-authored-by: karol-brejna-i <karolbrejna@apache.org>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

8 participants