Re-set default n_batch #17594
Merged
Re-set default n_batch #17594
+1
−1
Conversation
This file contains 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
dosubot
bot
added
the
size:XS
|
The latest updates on your projects. Learn more about Vercel for Git ↗︎ 1 Ignored Deployment
|
dosubot
bot
added
Ɑ: embeddings
baskaryan
approved these changes
Mar 29, 2024
dosubot
bot
added
the
lgtm
gkorland
pushed a commit
to FalkorDB/langchain
that referenced
this pull request
Mar 30, 2024
…in-ai#17594) When testing Nomic embeddings -- ``` from langchain_community.embeddings import LlamaCppEmbeddings embd_model_path = "/Users/rlm/Desktop/Code/llama.cpp/models/nomic-embd/nomic-embed-text-v1.Q4_K_S.gguf" embd_lc = LlamaCppEmbeddings(model_path=embd_model_path) embedding_lc = embd_lc.embed_query(query) ``` We were seeing this error for strings > a certain size -- ``` File ~/miniforge3/envs/llama2/lib/python3.9/site-packages/llama_cpp/llama.py:827, in Llama.embed(self, input, normalize, truncate, return_count) 824 s_sizes = [] 826 # add to batch --> 827 self._batch.add_sequence(tokens, len(s_sizes), False) 828 t_batch += n_tokens 829 s_sizes.append(n_tokens) File ~/miniforge3/envs/llama2/lib/python3.9/site-packages/llama_cpp/_internals.py:542, in _LlamaBatch.add_sequence(self, batch, seq_id, logits_all) 540 self.batch.token[j] = batch[i] 541 self.batch.pos[j] = i --> 542 self.batch.seq_id[j][0] = seq_id 543 self.batch.n_seq_id[j] = 1 544 self.batch.logits[j] = logits_all ValueError: NULL pointer access ``` The default `n_batch` of llama-cpp-python's Llama is `512` but we were explicitly setting it to `8`. These need to be set to equal for embedding models. * The embedding.cpp example has an assertion to make sure these are always equal. * Apparently this is not being done properly in llama-cpp-python. With `n_batch` set to 8, if more than 8 tokens are passed the batch runs out of space and it crashes. This also explains why the CPU compute buffer size was small: raw client with default `n_batch=512` ``` llama_new_context_with_model: CPU input buffer size = 3.51 MiB llama_new_context_with_model: CPU compute buffer size = 21.00 MiB ``` langchain with `n_batch=8` ``` llama_new_context_with_model: CPU input buffer size = 0.04 MiB llama_new_context_with_model: CPU compute buffer size = 0.33 MiB ``` We can work around this by passing `n_batch=512`, but this will not be obvious to some users: ``` embedding = LlamaCppEmbeddings(model_path=embd_model_path, n_batch=512) ``` From discussion w/ @cebtenzzre. Related: abetlen/llama-cpp-python#1189 Co-authored-by: Bagatur <[email protected]>
hinthornw
pushed a commit
that referenced
this pull request
Apr 26, 2024
When testing Nomic embeddings --
```
from langchain_community.embeddings import LlamaCppEmbeddings
embd_model_path = "/Users/rlm/Desktop/Code/llama.cpp/models/nomic-embd/nomic-embed-text-v1.Q4_K_S.gguf"
embd_lc = LlamaCppEmbeddings(model_path=embd_model_path)
embedding_lc = embd_lc.embed_query(query)
```
We were seeing this error for strings > a certain size --
```
File ~/miniforge3/envs/llama2/lib/python3.9/site-packages/llama_cpp/llama.py:827, in Llama.embed(self, input, normalize, truncate, return_count)
824 s_sizes = []
826 # add to batch
--> 827 self._batch.add_sequence(tokens, len(s_sizes), False)
828 t_batch += n_tokens
829 s_sizes.append(n_tokens)
File ~/miniforge3/envs/llama2/lib/python3.9/site-packages/llama_cpp/_internals.py:542, in _LlamaBatch.add_sequence(self, batch, seq_id, logits_all)
540 self.batch.token[j] = batch[i]
541 self.batch.pos[j] = i
--> 542 self.batch.seq_id[j][0] = seq_id
543 self.batch.n_seq_id[j] = 1
544 self.batch.logits[j] = logits_all
ValueError: NULL pointer access
```
The default `n_batch` of llama-cpp-python's Llama is `512` but we were
explicitly setting it to `8`.
These need to be set to equal for embedding models.
* The embedding.cpp example has an assertion to make sure these are
always equal.
* Apparently this is not being done properly in llama-cpp-python.
With `n_batch` set to 8, if more than 8 tokens are passed the batch runs
out of space and it crashes.
This also explains why the CPU compute buffer size was small:
raw client with default `n_batch=512`
```
llama_new_context_with_model: CPU input buffer size = 3.51 MiB
llama_new_context_with_model: CPU compute buffer size = 21.00 MiB
```
langchain with `n_batch=8`
```
llama_new_context_with_model: CPU input buffer size = 0.04 MiB
llama_new_context_with_model: CPU compute buffer size = 0.33 MiB
```
We can work around this by passing `n_batch=512`, but this will not be
obvious to some users:
```
embedding = LlamaCppEmbeddings(model_path=embd_model_path,
n_batch=512)
```
From discussion w/ @cebtenzzre. Related:
abetlen/llama-cpp-python#1189
Co-authored-by: Bagatur <[email protected]>
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.
When testing Nomic embeddings --
We were seeing this error for strings > a certain size --
The default
n_batchof llama-cpp-python's Llama is512but we were explicitly setting it to8.These need to be set to equal for embedding models.
With
n_batchset to 8, if more than 8 tokens are passed the batch runs out of space and it crashes.This also explains why the CPU compute buffer size was small:
raw client with default
n_batch=512langchain with
n_batch=8We can work around this by passing
n_batch=512, but this will not be obvious to some users:From discussion w/ @cebtenzzre. Related:
abetlen/llama-cpp-python#1189