68455353f5
feat(launcher): add disable_custom_kernels arg ( #67 )
2023-02-15 16:23:45 +01:00
c5a4a1faf3
feat(server): improve download logging ( #66 )
2023-02-15 16:11:32 +01:00
0fbc691946
feat: add safetensors conversion ( #63 )
2023-02-14 13:02:16 +01:00
9af454142a
feat: add distributed tracing ( #62 )
2023-02-13 13:02:45 +01:00
e520d5b349
fixed SSE naming ( #61 )
...
https://en.wikipedia.org/wiki/Server-sent_events
2023-02-08 22:30:11 +01:00
1ad3250b89
fix(docker): increase shm size ( #60 )
2023-02-08 17:53:33 +01:00
c503a639b1
feat(server): support t5 ( #59 )
2023-02-07 18:25:17 +01:00
2fe5e1b30e
V0.2.1 ( #58 )
2023-02-07 15:40:25 +01:00
4acc42a605
fix(server): better handling of inference mode ( #57 )
2023-02-07 15:38:22 +01:00
e114d87486
feat(ci): push to AML registry ( #56 )
2023-02-06 14:33:56 +01:00
a0dca443dd
feat(docs): Clarify installation steps ( #54 )
...
Adds some bits for first-time users (like me 😄 )
2023-02-03 13:07:55 +01:00
20c3c5940c
feat(router): refactor API and add openAPI schemas ( #53 )
2023-02-03 12:43:37 +01:00
b1482d9048
breaking(router): modify /generate API to only return generated text ( #50 )
...
@njhill, @yk FYI
generated_text was concatenated to the user prompt for legacy reason. We
want to remove this behaviour as we don't think it is useful and even
detrimonial to usability.
We also remove the unused Vec.
2023-02-02 15:02:04 +01:00
7b870e1e18
feat(router): use background task to manage request queue ( #52 )
...
Co-authored-by: Nick Hill <nickhill@us.ibm.com>
2023-02-02 14:59:27 +01:00
df227ac20d
fix(server): allow greedy repetition penalty ( #51 )
2023-02-02 10:34:35 +01:00
775115e3a5
feat(server): allow the server to use a local weight cache ( #49 )
2023-02-01 16:22:10 +01:00
313194f6d7
feat(server): support repetition penalty ( #47 )
2023-02-01 15:58:42 +01:00
2ad895a6cc
feat(server): allow gpt-neox models with odd vocab sizes to be sharded ( #48 )
2023-02-01 14:43:59 +01:00
404ed7a1f6
feat(ci): Docker build and push ( #46 )
2023-01-31 20:14:05 +01:00
f830706b21
feat(server): Support GPT-Neox ( #39 )
2023-01-31 18:53:56 +01:00
c6e8b9442b
fix(server): fix quantization for sharded models ( #45 )
2023-01-31 17:40:38 +01:00
017a2a8c2f
feat: Add token streaming using ServerSideEvents support ( #41 )
2023-01-31 17:04:00 +01:00
54fec93193
fix(server): fix seeding with multiple shards ( #44 )
2023-01-31 16:01:15 +01:00
03bdf18290
fix(server): fix seeding on gpu ( #42 )
2023-01-31 14:30:33 +01:00
4f9ac67cfa
Revert "feat: Add token streaming using ServerSideEvents support" ( #40 )
...
Reverts huggingface/text-generation-inference#36
2023-01-31 14:21:51 +01:00
7fbfbb0dc5
feat: Add token streaming using ServerSideEvents support ( #36 )
...
Add token streaming using ServerSideEvents (SSE).
The signature of the SSE events is:
```rust
struct Details {
finish_reason: String,
generated_tokens: u32,
seed: Option<u64>,
}
struct StreamResponse {
token: Token,
generated_text: Option<String>,
details: Option<Details>,
}
struct ErrorResponse {
error: String,
}
```
2023-01-31 11:49:43 +01:00
cd298bc5e5
feat: Support sampling seeding ( #37 )
...
Co-authored-by: Yannic Kilcher <yk@users.noreply.github.com>
2023-01-30 15:36:16 +01:00
1539d3cbbe
feat(router): Remove second lock from batcher hot path ( #27 )
...
@njhill
2023-01-26 16:29:13 +01:00
ce960be0a5
feat(bloom): use torch.nn.Linear and torch.nn.GELU ( #33 )
2023-01-26 15:33:45 +01:00
13e7044ab7
fix(dockerfile): fix docker build ( #32 )
2023-01-24 19:52:39 +01:00
5c01e2544c
fix(router): fix api-inference deployment ( #31 )
2023-01-23 17:42:14 +01:00
ab2ad91da3
fix(docker): fix api-inference deployment ( #30 )
2023-01-23 17:33:08 +01:00
f9d0ec376a
feat(docker): Make the image compatible with api-inference ( #29 )
2023-01-23 17:11:27 +01:00
1f570d181f
fix(server): Fix position ids ( #28 )
2023-01-20 15:35:22 +01:00
15511edc01
feat(server): Support SantaCoder ( #26 )
2023-01-20 12:24:39 +01:00
f7ac394935
fix(router): Obey max batch size ( #23 )
2023-01-17 09:11:21 +01:00
e6d3eb5d5d
fix(server): Minor refactorization using new_zeros ( #24 )
...
- Fix some type hints, in particular base tokenizer class
- Make use of `tensor.new_zero/empty` methods
- Simplify env var string parsing in launcher
2023-01-17 09:10:22 +01:00
fcc2c5fcbf
feat(launcher): Log server stdout ( #19 )
...
Co-authored-by: Nick Hill <nickhill@us.ibm.com>
2023-01-05 12:01:23 +01:00
b94f30215f
fix(server): Use cleanup_tokenization_spaces=False for lossless decoding ( #13 )
...
Fixes #12 in the easiest way I could think of.
2023-01-03 11:07:05 +01:00
60472f9d2b
feat(router): Add const parameters to validation logic ( #15 )
...
I noticed some opportunity to collapse some of the logic, in case you
are interested.
2023-01-03 10:41:22 +01:00
3efa5bbbfd
fix(router): Include special tokens when tokenizing ( #14 )
...
There's currently a discrepancy in the tokenization between the router
and python server code. The latter includes special tokens but former
does not.
This results in a token count mismatch for seq2seq models such as mt0
where the tokenizer emits an EOS token at the end.
This in turn results in some unexpected/incorrect output, in particular
when batch concatenation is involved, because the python code uses the
input length passed from the router for each row.
As far as I can tell, it is better to include this token in the encoder
`input_ids`, so I guess it's best to just adjust on the router side.
2022-12-30 19:31:44 +01:00
686cc66717
fix(server): Check for device type correctly when determining initial padding ( #16 )
...
AFAIK there is no torch device type called "gpu".
2022-12-30 19:30:42 +01:00
611e21cb13
fix(server): Fix stop sequences ( #11 )
2022-12-16 16:03:39 +01:00
3e2e6240b8
feat(launcher): Add integration tests ( #9 )
2022-12-16 11:29:36 +01:00
32a253063d
feat: Return logprobs ( #8 )
2022-12-15 17:03:56 +01:00
718096f695
feat: Support stop sequences ( #7 )
2022-12-12 18:25:22 +01:00
042180d88f
fix(server): Only pad to multiple of 8 on GPUs
2022-12-08 19:37:37 +01:00
a2985036aa
feat(server): Add model tests ( #6 )
2022-12-08 18:49:33 +01:00
31d76e238d
fix(batching): Avoid theoretical hang in batcher loop ( #5 )
...
- Avoid theoretical hang in batcher loop
- Avoid a couple of clones in the router generate method
- Keep attention mask tensors as integers
- Remove num_heads attribute
Co-authored-by: OlivierDehaene <Olivier.dehaene@gmail.com>
2022-12-05 10:10:59 +01:00
daa1d81d5e
feat(server): Support Galactica ( #4 )
2022-12-01 19:31:54 +01:00
OlivierDehaene