lalamo 0.5.6__py3-none-any.whl → 0.5.7__py3-none-any.whl

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (9) hide show
  1. lalamo/__init__.py +1 -1
  2. lalamo/speculator/estimator.py +2 -1
  3. lalamo/speculator/inference.py +3 -1
  4. {lalamo-0.5.6.dist-info → lalamo-0.5.7.dist-info}/METADATA +1 -1
  5. {lalamo-0.5.6.dist-info → lalamo-0.5.7.dist-info}/RECORD +9 -9
  6. {lalamo-0.5.6.dist-info → lalamo-0.5.7.dist-info}/WHEEL +0 -0
  7. {lalamo-0.5.6.dist-info → lalamo-0.5.7.dist-info}/entry_points.txt +0 -0
  8. {lalamo-0.5.6.dist-info → lalamo-0.5.7.dist-info}/licenses/LICENSE +0 -0
  9. {lalamo-0.5.6.dist-info → lalamo-0.5.7.dist-info}/top_level.txt +0 -0
lalamo/__init__.py CHANGED
@@ -15,7 +15,7 @@ from lalamo.speculator import (
15
15
  SpeculatorTrainingEvent,
16
16
  )
17
17
 
18
- __version__ = "0.5.6"
18
+ __version__ = "0.5.7"
19
19
 
20
20
  __all__ = [
21
21
  "AssistantMessage",
lalamo/speculator/estimator.py CHANGED
@@ -19,13 +19,14 @@ def estimate_memory_from_batchsize(
19
19
  memory_analysis = (
20
20
  jax.jit(
21
21
  functools.partial(
22
- model.generate_tokens,
22
+ LanguageModel.generate_tokens,
23
23
  max_output_length=max_output_length,
24
24
  num_top_logits_to_return=num_logits_per_token,
25
25
  ),
26
26
  backend="cpu", # cuda backend tries to allocate in .compile() and ooms
27
27
  )
28
28
  .lower(
29
+ model,
29
30
  prompt_token_ids=jax.ShapeDtypeStruct((batch_size, max_input_length), jnp.int32),
30
31
  prompt_lengths_without_padding=jax.ShapeDtypeStruct((batch_size,), jnp.int32),
31
32
  )
lalamo/speculator/inference.py CHANGED
@@ -30,12 +30,13 @@ def inference_collect_traces(
30
30
  generate_tokens_compiled = (
31
31
  jax.jit(
32
32
  functools.partial(
33
- model.generate_tokens,
33
+ LanguageModel.generate_tokens,
34
34
  max_output_length=max_output_length,
35
35
  num_top_logits_to_return=num_top_logits_to_collect,
36
36
  ),
37
37
  )
38
38
  .lower(
39
+ model,
39
40
  prompt_token_ids=jax.ShapeDtypeStruct((batch_size, max_input_length), jnp.int32),
40
41
  prompt_lengths_without_padding=jax.ShapeDtypeStruct((batch_size,), jnp.int32),
41
42
  )
@@ -60,6 +61,7 @@ def inference_collect_traces(
60
61
  )
61
62
 
62
63
  generated = generate_tokens_compiled(
64
+ model,
63
65
  prompt_token_ids=padded,
64
66
  prompt_lengths_without_padding=length_without_padding,
65
67
  )
{lalamo-0.5.6.dist-info → lalamo-0.5.7.dist-info}/METADATA RENAMED
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: lalamo
3
- Version: 0.5.6
3
+ Version: 0.5.7
4
4
  Summary: JAX library for optimization and export of models for use with the UZU inference engine.
5
5
  Requires-Python: <4,>=3.12
6
6
  Description-Content-Type: text/markdown
{lalamo-0.5.6.dist-info → lalamo-0.5.7.dist-info}/RECORD RENAMED
@@ -1,4 +1,4 @@
1
- lalamo/__init__.py,sha256=fm8ex5RjIqgZu_yH0WuOvZA1vN4EzqaDwuHThoOT35M,762
1
+ lalamo/__init__.py,sha256=ynqoSW4I6eQ92AR9jzr_8XMEiGpDGjMHHGUrjlTkmb0,762
2
2
  lalamo/common.py,sha256=5NUFD26yQgOnEEk3LaQnce8n-VwJxILkEpFesHZhtQU,3820
3
3
  lalamo/main.py,sha256=Lqv-lU6hRSqbJeVOheZoKllK1LaPHTuR_8jNTPC7UZg,23956
4
4
  lalamo/message_processor.py,sha256=bSUAQg7CemLTnBV4LtPxJBicAalruDCA-JXjkTYPZ8U,5797
@@ -76,13 +76,13 @@ lalamo/modules/token_mixers/state/kv_cache.py,sha256=QfnS3XgSmyDI9MBUbeLI4ABHLxi
76
76
  lalamo/modules/token_mixers/state/mamba_state.py,sha256=LHzJvNE6MkB7nrsZSNto6pxbnMJCl--JOoe9Fkcc9Mg,1642
77
77
  lalamo/speculator/__init__.py,sha256=9-tmZcbCom_lIGpJYn6xLlnEahFLFidpqmgkafmu--k,456
78
78
  lalamo/speculator/common.py,sha256=PudF_gkpe5_nQ-57sAC-foE1xCy_H2Axh5KwRoA86lo,587
79
- lalamo/speculator/estimator.py,sha256=SFF24TtxGVbb3hEziZ40A7aNdBd6_ZZJSX-CqmPhvA8,2503
80
- lalamo/speculator/inference.py,sha256=EzNLmdP-AOtIN8C39-fq-3ZnKX_yxOOwMufDEmi54zM,3624
79
+ lalamo/speculator/estimator.py,sha256=4D8dPZCWsrpORb7y8pQ6VsiIg1Cblvvxe6gXCoYtcD4,2530
80
+ lalamo/speculator/inference.py,sha256=5GntUgj0HQLeLn3HIHnVX8EEO0EBzmKeP5-_U7kdFAM,3670
81
81
  lalamo/speculator/ngram.py,sha256=95mdfAWhx4d5XOnOwhyhElnvcy6nlUjYhcbJzqDs414,5875
82
82
  lalamo/speculator/utils.py,sha256=0wZoMMIzzk0Q-3zq5H5f-JBplePNHxywndkrNtOJOyo,1697
83
- lalamo-0.5.6.dist-info/licenses/LICENSE,sha256=diHRfjSEJHD1nnEeMIfMRCjR3UERf8bT3eseD6b1ayA,1072
84
- lalamo-0.5.6.dist-info/METADATA,sha256=CxuLNGMlIPufGa9ulMtiVsduvt_Kr01706jdFHdiDwk,3146
85
- lalamo-0.5.6.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
86
- lalamo-0.5.6.dist-info/entry_points.txt,sha256=qli7qTfnBk5WP10rOGXXEckHMtt-atJMDWd8jN89Uks,43
87
- lalamo-0.5.6.dist-info/top_level.txt,sha256=VHvWL5JN5XRG36NsN_MieJ7EwRihEOrEjyDaTdFJ-aI,7
88
- lalamo-0.5.6.dist-info/RECORD,,
83
+ lalamo-0.5.7.dist-info/licenses/LICENSE,sha256=diHRfjSEJHD1nnEeMIfMRCjR3UERf8bT3eseD6b1ayA,1072
84
+ lalamo-0.5.7.dist-info/METADATA,sha256=ZzSQiR7KYoAXkBKOToa2bi9tCPRvVXvhnN5y6AO7wyQ,3146
85
+ lalamo-0.5.7.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
86
+ lalamo-0.5.7.dist-info/entry_points.txt,sha256=qli7qTfnBk5WP10rOGXXEckHMtt-atJMDWd8jN89Uks,43
87
+ lalamo-0.5.7.dist-info/top_level.txt,sha256=VHvWL5JN5XRG36NsN_MieJ7EwRihEOrEjyDaTdFJ-aI,7
88
+ lalamo-0.5.7.dist-info/RECORD,,
{lalamo-0.5.6.dist-info → lalamo-0.5.7.dist-info}/WHEEL RENAMED
File without changes