whisper : add batched decoding #1486
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
This should be ready to merge - will do so tomorrow |
During testing, I encountered a specific error which exhibits a 100% reproducibility rate. Audio: L'Océan et l'Humanité, destins liés ! _ Lamya Essemlali _ TEDxOrléans (320 kbps).zip |
|
@bobqianic Thanks for the feedback. If increase the factor to Edit: I just noticed you weren't using the latest version of this branch, so you had |
Sorry for the late reply, I was busy with some lab work this morning. Just checked, and yeah, you're right. The test version I used had factor = 2. Just now, I tested 4c245ea on both Windows and Linux. Looks like there's still something not quite right. Like, the transcription runs fine on Windows but the output looks weird. See this |
|
Are you using large v2? |
Yes |
|
Which command are you using? It works on my end: WHISPER_CUBLAS=1 make -j && ./main -m ./models-mnt/ggml-large-v2.bin -f ./samples/fr0.wav -bs 5 -t 6 -l auto |
|
Try 270b1e4 and let me know if the issue is resolved. |
It works! Edit: The strange characters mentioned earlier are likely caused by Windows not setting the terminal to UTF-8.
|
felrock
pushed a commit
to felrock/whisper.cpp
that referenced
this pull request
Nov 18, 2023
* whisper : add whisper_batch * whisper : move kv_self to whisper_state * whisper : full batched decoding support * whisper : fix memory leak in whisper_batch * whisper : fix mem leak again + remove oboslete function * whisper : clear kv cache when using whisper_decode API * whisper : speed-up sampling * whisper : fix decoders initializer * bench : add batch size 5 bench * whisper : add comment about the KV cache size * whisper : add check for max number of decoders * whisper : avoid starting sampling threads with bs=1 * whisper : enable beam-search by default * cuda : sync llama.cpp fixes
landtanin
pushed a commit
to landtanin/whisper.cpp
that referenced
this pull request
Dec 16, 2023
* whisper : add whisper_batch * whisper : move kv_self to whisper_state * whisper : full batched decoding support * whisper : fix memory leak in whisper_batch * whisper : fix mem leak again + remove oboslete function * whisper : clear kv cache when using whisper_decode API * whisper : speed-up sampling * whisper : fix decoders initializer * bench : add batch size 5 bench * whisper : add comment about the KV cache size * whisper : add check for max number of decoders * whisper : avoid starting sampling threads with bs=1 * whisper : enable beam-search by default * cuda : sync llama.cpp fixes
josharian
added a commit
to josharian/whisper.cpp
that referenced
this pull request
Mar 10, 2024
As of ggml-org#1486, whisper.cpp uses a unified KV cache with KQ masking. As a result, depending on their location in the batch, identical sequences in a batch can have slightly different outputs due to floating point rounding errors during reduction. See the discussion in ggml-org#1941 for more details. The beam search code used "has identical sum of log probabilities" as a shorthand for "is an identical token sequence". However, per above, identical tokens do not necessarily result in identical probabilities. Instead, explicitly compare on sequences. This is linear in cost when they are identical, but the lengths are always small and the comparisons are cheap. This increases diversity during beam search. This improves output quality for some short samples I've been working with, at no detectable performance cost. I haven't checked against larger corpuses. Fixes ggml-org#1941
ggerganov
pushed a commit
that referenced
this pull request
Mar 10, 2024
As of #1486, whisper.cpp uses a unified KV cache with KQ masking. As a result, depending on their location in the batch, identical sequences in a batch can have slightly different outputs due to floating point rounding errors during reduction. See the discussion in #1941 for more details. The beam search code used "has identical sum of log probabilities" as a shorthand for "is an identical token sequence". However, per above, identical tokens do not necessarily result in identical probabilities. Instead, explicitly compare on sequences. This is linear in cost when they are identical, but the lengths are always small and the comparisons are cheap. This increases diversity during beam search. This improves output quality for some short samples I've been working with, at no detectable performance cost. I haven't checked against larger corpuses. Fixes #1941
jiahansu
pushed a commit
to WiseSync/whisper.cpp
that referenced
this pull request
Apr 17, 2024
As of ggml-org#1486, whisper.cpp uses a unified KV cache with KQ masking. As a result, depending on their location in the batch, identical sequences in a batch can have slightly different outputs due to floating point rounding errors during reduction. See the discussion in ggml-org#1941 for more details. The beam search code used "has identical sum of log probabilities" as a shorthand for "is an identical token sequence". However, per above, identical tokens do not necessarily result in identical probabilities. Instead, explicitly compare on sequences. This is linear in cost when they are identical, but the lengths are always small and the comparisons are cheap. This increases diversity during beam search. This improves output quality for some short samples I've been working with, at no detectable performance cost. I haven't checked against larger corpuses. Fixes ggml-org#1941
viktor-silakov
pushed a commit
to viktor-silakov/whisper_node_mic.cpp
that referenced
this pull request
May 11, 2024
As of ggml-org#1486, whisper.cpp uses a unified KV cache with KQ masking. As a result, depending on their location in the batch, identical sequences in a batch can have slightly different outputs due to floating point rounding errors during reduction. See the discussion in ggml-org#1941 for more details. The beam search code used "has identical sum of log probabilities" as a shorthand for "is an identical token sequence". However, per above, identical tokens do not necessarily result in identical probabilities. Instead, explicitly compare on sequences. This is linear in cost when they are identical, but the lengths are always small and the comparisons are cheap. This increases diversity during beam search. This improves output quality for some short samples I've been working with, at no detectable performance cost. I haven't checked against larger corpuses. Fixes ggml-org#1941
iThalay
pushed a commit
to iThalay/whisper.cpp
that referenced
this pull request
Sep 23, 2024
* whisper : add whisper_batch * whisper : move kv_self to whisper_state * whisper : full batched decoding support * whisper : fix memory leak in whisper_batch * whisper : fix mem leak again + remove oboslete function * whisper : clear kv cache when using whisper_decode API * whisper : speed-up sampling * whisper : fix decoders initializer * bench : add batch size 5 bench * whisper : add comment about the KV cache size * whisper : add check for max number of decoders * whisper : avoid starting sampling threads with bs=1 * whisper : enable beam-search by default * cuda : sync llama.cpp fixes
iThalay
pushed a commit
to iThalay/whisper.cpp
that referenced
this pull request
Sep 23, 2024
As of ggml-org#1486, whisper.cpp uses a unified KV cache with KQ masking. As a result, depending on their location in the batch, identical sequences in a batch can have slightly different outputs due to floating point rounding errors during reduction. See the discussion in ggml-org#1941 for more details. The beam search code used "has identical sum of log probabilities" as a shorthand for "is an identical token sequence". However, per above, identical tokens do not necessarily result in identical probabilities. Instead, explicitly compare on sequences. This is linear in cost when they are identical, but the lengths are always small and the comparisons are cheap. This increases diversity during beam search. This improves output quality for some short samples I've been working with, at no detectable performance cost. I haven't checked against larger corpuses. Fixes ggml-org#1941
iThalay
pushed a commit
to iThalay/whisper.cpp
that referenced
this pull request
Sep 23, 2024
As of ggml-org#1486, whisper.cpp uses a unified KV cache with KQ masking. As a result, depending on their location in the batch, identical sequences in a batch can have slightly different outputs due to floating point rounding errors during reduction. See the discussion in ggml-org#1941 for more details. The beam search code used "has identical sum of log probabilities" as a shorthand for "is an identical token sequence". However, per above, identical tokens do not necessarily result in identical probabilities. Instead, explicitly compare on sequences. This is linear in cost when they are identical, but the lengths are always small and the comparisons are cheap. This increases diversity during beam search. This improves output quality for some short samples I've been working with, at no detectable performance cost. I haven't checked against larger corpuses. Fixes ggml-org#1941
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
ref #1048
Description
This PR implements efficient batched decoding. With CUDA, the speed with 5 beams is the same as with 1 beam so there is likely no reason to ever use 1 beam. With Metal, using more than 1 beam results in some slowdown since the Metal kernels do not scale as well with the batch size. Still, it is much faster compared to what was on
masterand the improved transcription quality might be worth it.Also, this PR:
CUDA_VISIBLE_DEVICES=-10.2finstead of0.4fTests
V100 bs=1 `master`
V100 bs=5 `master`
V100 bs=5 `batched`
In this case, we observe beam search with 5 beams is the same speed as 1 beam
Benches