Commit Graph

90 Commits

Author SHA1 Message Date
Jacek Czaja
5df20f88ff
Fix to non-LLAMA models (#177)
Co-authored-by: Jacek Czaja <jczaja@habana.ai>
2024-07-04 13:42:24 +02:00
Jacek Czaja
c64b5b75e2
[TORCH COMPILE] Ignore HPU GRAPHS env var when eager mode is used (#165)
Co-authored-by: Jacek Czaja <jczaja@habana.ai>
2024-07-03 15:17:27 +02:00
Jacek Czaja
ef86232c94
[Torch.compile] Enable llama-2-7b (#157)
Co-authored-by: Jacek Czaja <jczaja@habana.ai>
2024-06-14 15:56:23 +02:00
Karol Damaszke
0e8f8726db
Warmup all decode buckets (#152)
Co-authored-by: Karol Damaszke <kdamaszke@habana.ai>
2024-05-29 22:46:55 +02:00
Karol Damaszke
7b879fd1d8
Pad next token chooser parameters with empty logits processors (#151)
Co-authored-by: Karol Damaszke <kdamaszke@habana.ai>
2024-05-29 22:43:56 +02:00
Jimin Ha
1023de8048
Add flash_attention argument options for Mistral (#145)
Co-authored-by: Karol Damaszke <karol.damaszke@intel.com>
Co-authored-by: Karol Damaszke <kdamaszke@habana.ai>
2024-05-27 20:00:42 +02:00
Karol Damaszke
32acdd55b4
Add grammar support (#140)
Co-authored-by: Karol Damaszke <kdamaszke@habana.ai>
2024-05-20 11:16:34 +02:00
Karol Damaszke
bad7fe720a
Fix warmup shapes for corner cases (#136)
Co-authored-by: Karol Damaszke <kdamaszke@habana.ai>
2024-05-06 11:35:27 +02:00
Karol Damaszke
600d033c04 Merge branch 'habana-main' into rebase_tgi_2.0 2024-04-29 09:44:45 +03:00
regisss
37aabf8571
Move call to adapt_transformers_to_gaudi earlier in the code (#133) 2024-04-26 11:07:27 +02:00
drbh
e259625b8b fix: Handle concurrent grammar requests (#1610)
This PR fixes parallel grammar requests, currently grammar states are
not concatenated correctly when a new request is added to the batch and
this results in incorrect generation. This PR updates the `concatenate`
function to correctly include the previous states.

fixes: #1601
2024-04-25 10:11:40 +03:00
OlivierDehaene
666cdaaf16 feat: Qwen2 (#1608)
See #1584

---------

Co-authored-by: Cheng Kuan Yong Jason <jasoncky96@gmail.com>
2024-04-25 09:21:22 +03:00
Nicolas Patry
21d52c9ca1 Revamp medusa implementation so that every model can benefit. (#1588)
<!--
Congratulations! You've made it this far! You're not quite done yet
though.

Once merged, your PR is going to appear in the release notes with the
title you set, so make sure it's a great title that fully reflects the
extent of your awesome contribution.

Then, please replace this with a description of the change and which
issue is fixed (if applicable). Please also include relevant motivation
and context. List any dependencies (if any) that are required for this
change.

Once you're done, someone will review your PR shortly (see the section
"Who can review?" below to tag some potential reviewers). They may
suggest changes to make the code even better. If no one reviewed your PR
after a week has passed, don't hesitate to post a new comment
@-mentioning the same persons---sometimes notifications get lost.
-->

<!-- Remove if not applicable -->

Fixes # (issue)

- [ ] This PR fixes a typo or improves the docs (you can dismiss the
other checks if that's the case).
- [ ] Did you read the [contributor
guideline](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md#start-contributing-pull-requests),
      Pull Request section?
- [ ] Was this discussed/approved via a Github issue or the
[forum](https://discuss.huggingface.co/)? Please add a link
      to it if that's the case.
- [ ] Did you make sure to update the documentation with your changes?
Here are the
[documentation
guidelines](https://github.com/huggingface/transformers/tree/main/docs),
and
[here are tips on formatting
docstrings](https://github.com/huggingface/transformers/tree/main/docs#writing-source-documentation).
- [ ] Did you write any new necessary tests?

Anyone in the community is free to review the PR once the tests have
passed. Feel free to tag
members/contributors who may be interested in your PR.

<!-- Your PR will be replied to more quickly if you can figure out the
right person to tag with @

@OlivierDehaene OR @Narsil

 -->
2024-04-25 09:13:03 +03:00
drbh
55acb86f42 Outlines guided generation (#1539)
This WIP PR starts to add grammar support via outlines, currently this
PR supports very simple regex grammars and does not optimize for
precompiling or caching grammar fsm's.

todo:
- [X] add simple outlines guidance to `NextTokenChooser`
- [X] update protos for grammar
- [X] update generation params API
- [X] constrain simple grammar
- [ ] support parsing more complex grammar into fsm
- [ ] support all outline support grammar types
- [ ] explore optimizations to avoid recompiling grammars

guided request
```bash
curl -s 'http://localhost:3000/generate' \
--header 'Content-Type: application/json' \
--data-raw '{
    "inputs": "make an email for david: \n",
    "parameters": {
        "max_new_tokens": 6,
        "grammar": "[\\w-]+@([\\w-]+\\.)+[\\w-]+"
    }
}' | jq
```
response
```json
{
  "generated_text": "david@example.com"
}
```

unguided request
```bash
curl -s 'http://localhost:3000/generate' \
--header 'Content-Type: application/json' \
--data '{
    "inputs": "make an email for david: \n",
    "parameters": {
        "max_new_tokens": 6
    }
}' | jq
```
response
```json
{
  "generated_text": "    email = 'david"
}
```
2024-04-24 14:57:37 +03:00
OlivierDehaene
f1d8da3ba6 feat(server): add frequency penalty (#1541) 2024-04-24 08:43:50 +00:00
Nicolas Patry
433934519c Fixing top_n_tokens. (#1497)
Superseeds #1459

The fix works as follows.
We updated next_token_chooser to return all logprbs, then
batch_top_n_tokens, now also gets accepted_ids + speculated_length (so
it knows how to interpret the flat logprobs).

We then update the code to return lists ot `Tokens` that it expects.
<!--
Congratulations! You've made it this far! You're not quite done yet
though.

Once merged, your PR is going to appear in the release notes with the
title you set, so make sure it's a great title that fully reflects the
extent of your awesome contribution.

Then, please replace this with a description of the change and which
issue is fixed (if applicable). Please also include relevant motivation
and context. List any dependencies (if any) that are required for this
change.

Once you're done, someone will review your PR shortly (see the section
"Who can review?" below to tag some potential reviewers). They may
suggest changes to make the code even better. If no one reviewed your PR
after a week has passed, don't hesitate to post a new comment
@-mentioning the same persons---sometimes notifications get lost.
-->

<!-- Remove if not applicable -->

Fixes # (issue)

- [ ] This PR fixes a typo or improves the docs (you can dismiss the
other checks if that's the case).
- [ ] Did you read the [contributor
guideline](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md#start-contributing-pull-requests),
      Pull Request section?
- [ ] Was this discussed/approved via a Github issue or the
[forum](https://discuss.huggingface.co/)? Please add a link
      to it if that's the case.
- [ ] Did you make sure to update the documentation with your changes?
Here are the
[documentation
guidelines](https://github.com/huggingface/transformers/tree/main/docs),
and
[here are tips on formatting
docstrings](https://github.com/huggingface/transformers/tree/main/docs#writing-source-documentation).
- [ ] Did you write any new necessary tests?

Anyone in the community is free to review the PR once the tests have
passed. Feel free to tag
members/contributors who may be interested in your PR.

<!-- Your PR will be replied to more quickly if you can figure out the
right person to tag with @

@OlivierDehaene OR @Narsil

 -->
2024-04-23 08:49:24 +03:00
OlivierDehaene
5c9ef069ed feat: add more latency metrics in forward (#1346) 2024-04-19 13:41:34 +03:00
OlivierDehaene
79f268f95a chore: formatting 2024-04-18 16:26:00 +03:00
Nicolas Patry
a7f52f3812 Speculative (#1308) 2024-04-18 12:39:39 +00:00
Karol Damaszke
30cc78773e
Skip server tests of not enabled models (#125)
Co-authored-by: Karol Damaszke <kdamaszke@habana.ai>
2024-04-09 14:15:41 +02:00
Karol Damaszke
d957e32601
Add Habana copyright header (#122)
Co-authored-by: Karol Damaszke <kdamaszke@habana.ai>
2024-04-08 18:06:21 +02:00
Karol Damaszke
b0de25a285
Don't set rope_scaling for unsupported models (#115)
Co-authored-by: Karol Damaszke <kdamaszke@habana.ai>
2024-04-02 12:12:02 +02:00
Karol Damaszke
7342baa2eb
Add support for rope_scaling and remove is_optimized_for_gaudi (#112)
Co-authored-by: Karol Damaszke <kdamaszke@habana.ai>
2024-03-29 15:07:32 +01:00
Karol Damaszke
bf5263b88b
Disable watermark with FP8 quantization (#114)
Co-authored-by: Karol Damaszke <kdamaszke@habana.ai>
2024-03-27 13:32:20 +01:00
jkaniecki
56f00a552b
Adjust warmup to all possible bucket sizes and decode batch size = 1 (#113) 2024-03-27 11:59:51 +01:00
Karol Damaszke
b45f648483
Add warmup for logits processors (#107)
Co-authored-by: Karol Damaszke <kdamaszke@habana.ai>
2024-03-18 15:17:47 +01:00
yuanwu2017
a4d5c3f40f
Fix the generate_stream crash in concurrent query (#105)
Signed-off-by: yuanwu <yuan.wu@intel.com>
2024-03-15 10:54:56 +01:00
Karol Damaszke
80ae9ead28
Set MAX_TOTAL_TOKENS automatically (#91)
Co-authored-by: Karol Damaszke <kdamaszke@habana.ai>
2024-03-01 11:25:15 +01:00
Karol Damaszke
a5c788cfe4
Remove redundant fill op (#83) (#90)
Co-authored-by: mswiniarsk <156412439+mswiniarsk@users.noreply.github.com>
2024-03-01 01:32:02 +01:00
Karol Damaszke
03c2123244
Use batched index_copy (#73) (#89)
Co-authored-by: madamczykhabana <110973826+madamczykhabana@users.noreply.github.com>
2024-02-29 15:45:16 +01:00
Karol Damaszke
7dbf4bf7a4
Improve tensor slicing performance (#66) (#87)
Co-authored-by: mswiniarsk <156412439+mswiniarsk@users.noreply.github.com>
2024-02-29 10:48:54 +01:00
Karol Damaszke
3831f1bed5
Add warmup for shift operation (#59) (#86) 2024-02-29 09:19:28 +01:00
Karol Damaszke
022ce1eaaf
Overhead reduction (#58) (#85)
Co-authored-by: mrs303 <54661797+mrs303@users.noreply.github.com>
2024-02-29 09:17:45 +01:00
Karol Damaszke
212136dff8
Log exceptions to debug.log (#52) (#84)
Co-authored-by: madamczykhabana <110973826+madamczykhabana@users.noreply.github.com>
2024-02-29 09:14:42 +01:00
Karol Damaszke
c7ccfb87ff
Grouped pad/shift/move operations (#57) (#82)
Co-authored-by: madamczykhabana <110973826+madamczykhabana@users.noreply.github.com>
2024-02-29 04:16:44 +01:00
Karol Damaszke
2122acc60f
Add warmup for all possible shapes for prefill #49 (#81) 2024-02-28 10:40:13 +01:00
Karol Damaszke
31bed905d4
Update habana profiler (#50) (#80)
Co-authored-by: mswiniarsk <156412439+mswiniarsk@users.noreply.github.com>
2024-02-28 09:57:40 +01:00
Karol Damaszke
941d36f3fd
Enable deferred token generation (#44) (#75)
Co-authored-by: Krzysztof Laskowski <klaskowski@habana.ai>
2024-02-27 15:46:40 +01:00
jkaniecki
83b059bd27
Bulk shifting (#40) (#70)
Co-authored-by: madamczykhabana <110973826+madamczykhabana@users.noreply.github.com>
2024-02-26 17:29:56 +01:00
jkaniecki
c3bd8ef445
Add Fp8 support (#42) (#71)
Co-authored-by: mrs303 <54661797+mrs303@users.noreply.github.com>
Co-authored-by: Adam Stachowicz <105052242+astachowiczhabana@users.noreply.github.com>
Co-authored-by: Grzegorz Morys <gmorys@habana.ai>
2024-02-23 11:52:28 +01:00
jkaniecki
a490847702
Sequence bucketing for prefill (#39) (#67)
Co-authored-by: mswiniarsk <156412439+mswiniarsk@users.noreply.github.com>
2024-02-23 01:52:14 +01:00
jkaniecki
9ad6086250
Improve habana profile dev experience (#36) (#65)
Co-authored-by: Michal Szutenberg <37601244+szutenberg@users.noreply.github.com>
2024-02-22 13:57:45 +01:00
jkaniecki
8f590759e3
Prefill optimization by allocating space only for the first output token (#34) (#62)
Co-authored-by: mswiniarsk <156412439+mswiniarsk@users.noreply.github.com>
Co-authored-by: Karol Damaszke <karol.damaszke@intel.com>
2024-02-22 04:55:43 +01:00
jkaniecki
80303b469c
Do not limit hpu graphs by default (#32) (#61)
Co-authored-by: mswiniarsk <156412439+mswiniarsk@users.noreply.github.com>
2024-02-21 15:38:00 +01:00
Karol Damaszke
2a7a967de3
Revert prefill optimization and fix accuracy issue in shift operation (#29)
Co-authored-by: Karol Damaszke <kdamaszke@habana.ai>
Co-authored-by: madamczykhabana <110973826+madamczykhabana@users.noreply.github.com>
Co-authored-by: jkaniecki <153085639+jkaniecki@users.noreply.github.com>
2024-01-23 15:19:07 +01:00
jkaniecki
ac3bc0e95e
Removed kv_cache from HPU graph output (#19) 2024-01-19 15:34:13 +01:00
Karol Damaszke
60f63262db
Prefill optimization by allocating space only for the first token (#17) 2024-01-19 15:18:35 +01:00
Adam Stachowicz
0b96da89aa
Make tokenizer optional (#12) 2024-01-19 15:12:04 +01:00
madamczykhabana
381ec38cad
Batch bucketing improvements (#15) 2024-01-17 10:09:27 +01:00
mrs303
8523f7ef64
Deepspeed terminate (#11) 2024-01-17 09:57:03 +01:00