lalamo 0.6.2__py3-none-any.whl → 0.6.3__py3-none-any.whl

lalamo/__init__.py

@@ -1,3 +1,8 @@
+import os
+
+# Must run before importing jax / tensorflow, this hides the XLA optimization logs
+os.environ.setdefault("TF_CPP_MIN_LOG_LEVEL", "3")
+
 from lalamo.commands import (
     CollectTracesCallbacks,
     ConversionCallbacks,
@@ -27,7 +32,7 @@ from lalamo.speculator import (
     SpeculatorTrainingEvent,
 )
 
-__version__ = "0.6.2"
+__version__ = "0.6.3"
 
 __all__ = [
     "AssistantMessage",

lalamo/speculator/estimator.py

@@ -1,5 +1,6 @@
 import functools
 import itertools
+import os
 from collections.abc import Callable
 from typing import NamedTuple
 
@@ -10,10 +11,35 @@ from lalamo.models import LanguageModel
 
 
 def get_default_device_memory() -> int | None:
+    dynamic_allocate = False
+
+    preallocate = os.getenv("XLA_PYTHON_CLIENT_PREALLOCATE", "")
+    dynamic_allocate |= preallocate.strip().lower() in {"0", "false", "no", "off"}
+
+    allocator = os.getenv("XLA_PYTHON_CLIENT_ALLOCATOR", "")
+    dynamic_allocate |= allocator.strip().lower() in {"platform", "cuda_malloc_async"}
+
+    if dynamic_allocate:
+        return None
+
     memory_stats = jax.local_devices()[0].memory_stats()
     if memory_stats is None or "bytes_limit" not in memory_stats:
         return None
-    return memory_stats["bytes_limit"]
+
+    # discount by 0.98 because if you try to allocate exactly bytes_limit
+    # jax will _try_ to allocate a bit more, and then throws an OOM
+    bytes_limit = memory_stats["bytes_limit"] * 0.98
+
+    mem_fraction_raw = os.getenv("XLA_PYTHON_CLIENT_MEM_FRACTION", "")
+    try:
+        mem_fraction = float(mem_fraction_raw)
+    except ValueError:
+        mem_fraction = 0.75  # jax default https://docs.jax.dev/en/latest/gpu_memory_allocation.html
+
+    # JAX usually can't allocate more than 98%-ish percent memory; idk why
+    # Besides we use _some_ autotuning during runtime, so we add 0.96 safety margin here
+    bytes_limit = int(max(bytes_limit, bytes_limit / min(mem_fraction, 1.0)) * 0.96)
+    return bytes_limit
 
 
 def estimate_memory_from_batchsize(
@@ -30,14 +56,14 @@ def estimate_memory_from_batchsize(
                 max_output_length=max_output_length,
                 num_top_logits_to_return=num_logits_per_token,
             ),
-            backend="cpu",  # cuda backend tries to allocate in .compile() and ooms
         )
         .lower(
             model,
             prompt_token_ids=jax.ShapeDtypeStruct((batch_size, max_input_length), jnp.int32),
             prompt_lengths_without_padding=jax.ShapeDtypeStruct((batch_size,), jnp.int32),
         )
-        .compile()
+        # disables autotune, see https://guides.lw1.at/all-xla-options/#--xla_gpu_autotune_level
+        .compile(compiler_options={"xla_gpu_autotune_level": "0"})
         .memory_analysis()
     )
 

lalamo/speculator/inference.py

@@ -40,7 +40,12 @@ def inference_collect_traces(
             prompt_token_ids=jax.ShapeDtypeStruct((batch_size, max_input_length), jnp.int32),
             prompt_lengths_without_padding=jax.ShapeDtypeStruct((batch_size,), jnp.int32),
         )
-        .compile()
+        # the autotune levels are (according to https://guides.lw1.at/all-xla-options/#--xla_gpu_autotune_level)
+        # 0 - no autotune, gpu shouldn't be touched
+        # 1 - basic level, gpu should be touched veeery little
+        # 2,3 - gpu touched more and more
+        # 4 (default) - gpu might allocate more memory than the run would require!
+        .compile(compiler_options={"xla_gpu_autotune_level": "2"})
     )
 
     prefixes = chain.from_iterable(map(get_prefixes_ending_in_user_message, conversations))

{lalamo-0.6.2.dist-info → lalamo-0.6.3.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lalamo
-Version: 0.6.2
+Version: 0.6.3
 Summary: JAX library for optimization and export of models for use with the UZU inference engine.
 Requires-Python: <4,>=3.12
 Description-Content-Type: text/markdown

{lalamo-0.6.2.dist-info → lalamo-0.6.3.dist-info}/RECORD

@@ -1,4 +1,4 @@
-lalamo/__init__.py,sha256=LlHkzLyEJUp3M_Zvcwqc2CTzLhhyqnHGqk2bs6hV3fY,1386
+lalamo/__init__.py,sha256=MBRNUkooL0qiMz4B9VqvRblslw0VXAODjzg3BtY7uhU,1532
 lalamo/commands.py,sha256=zXyyrLTHhP9wouwtpX4RUZeEF6No-_9ee-y_GWGhw7k,10972
 lalamo/common.py,sha256=WaNJx20eUX4CBF50aym9lniGAiX-SzBJzDzO5Jh6zXA,4312
 lalamo/main.py,sha256=Tez84CtMxUi1ySuRSqQElu4Zr1UWs_Gw6HX1xtCZknQ,27383
@@ -83,13 +83,13 @@ lalamo/modules/token_mixers/state/mamba_state.py,sha256=LHzJvNE6MkB7nrsZSNto6pxb
 lalamo/modules/token_mixers/state/short_conv_state.py,sha256=osjcDHoeFWQaUoOROzeJe8F1qC8rvqunimGD4CuIDHo,895
 lalamo/speculator/__init__.py,sha256=9-tmZcbCom_lIGpJYn6xLlnEahFLFidpqmgkafmu--k,456
 lalamo/speculator/common.py,sha256=PudF_gkpe5_nQ-57sAC-foE1xCy_H2Axh5KwRoA86lo,587
-lalamo/speculator/estimator.py,sha256=S_TRwMnjWg5qt9le2AYua_Vmo6QkIT-0Si7TjCfC7xc,2670
-lalamo/speculator/inference.py,sha256=5GntUgj0HQLeLn3HIHnVX8EEO0EBzmKeP5-_U7kdFAM,3670
+lalamo/speculator/estimator.py,sha256=xkvrNsoIWUbotv__e2ms-QP3NUnj92W_Fcpse8rvQ_o,3797
+lalamo/speculator/inference.py,sha256=uEv33Qqcpa2xqEKdIzmPzkAzRsZOlb8TPeEG6TP6fjo,4071
 lalamo/speculator/ngram.py,sha256=2eqInIieJPaQHCvIfnCIDtwMa8PGEtiND_NkG7plE34,5899
 lalamo/speculator/utils.py,sha256=0wZoMMIzzk0Q-3zq5H5f-JBplePNHxywndkrNtOJOyo,1697
-lalamo-0.6.2.dist-info/licenses/LICENSE,sha256=diHRfjSEJHD1nnEeMIfMRCjR3UERf8bT3eseD6b1ayA,1072
-lalamo-0.6.2.dist-info/METADATA,sha256=ZxR_Z-Q90tm45WUk2Wh1e_SpjKT0oW-FvkqmNXAqdvA,3112
-lalamo-0.6.2.dist-info/WHEEL,sha256=wUyA8OaulRlbfwMtmQsvNngGrxQHAvkKcvRmdizlJi0,92
-lalamo-0.6.2.dist-info/entry_points.txt,sha256=qli7qTfnBk5WP10rOGXXEckHMtt-atJMDWd8jN89Uks,43
-lalamo-0.6.2.dist-info/top_level.txt,sha256=VHvWL5JN5XRG36NsN_MieJ7EwRihEOrEjyDaTdFJ-aI,7
-lalamo-0.6.2.dist-info/RECORD,,
+lalamo-0.6.3.dist-info/licenses/LICENSE,sha256=diHRfjSEJHD1nnEeMIfMRCjR3UERf8bT3eseD6b1ayA,1072
+lalamo-0.6.3.dist-info/METADATA,sha256=tvl435CXfVSezsey_FhFzZ5YaY64GOjZQBaDY5vEAPU,3112
+lalamo-0.6.3.dist-info/WHEEL,sha256=wUyA8OaulRlbfwMtmQsvNngGrxQHAvkKcvRmdizlJi0,92
+lalamo-0.6.3.dist-info/entry_points.txt,sha256=qli7qTfnBk5WP10rOGXXEckHMtt-atJMDWd8jN89Uks,43
+lalamo-0.6.3.dist-info/top_level.txt,sha256=VHvWL5JN5XRG36NsN_MieJ7EwRihEOrEjyDaTdFJ-aI,7
+lalamo-0.6.3.dist-info/RECORD,,