lalamo 0.6.2__tar.gz → 0.6.4__tar.gz

{lalamo-0.6.2 → lalamo-0.6.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lalamo
-Version: 0.6.2
+Version: 0.6.4
 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 → lalamo-0.6.4}/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__ = "0.6.4"
 
 __all__ = [
     "AssistantMessage",

{lalamo-0.6.2 → lalamo-0.6.4}/lalamo/main.py

@@ -49,7 +49,10 @@ from lalamo.message_processor import UserMessage
 from lalamo.model_import import REPO_TO_MODEL, ModelSpec
 from lalamo.model_import.common import FileSpec
 from lalamo.models import ClassifierModelConfig, LanguageModelConfig
-from lalamo.speculator.estimator import get_default_device_memory
+from lalamo.speculator.estimator import (
+    get_default_device_bytes,
+    get_usable_memory_from_bytes,
+)
 from lalamo.speculator.ngram import NGramSpeculator
 from lalamo.speculator.utils import test_speculator
 
@@ -384,6 +387,7 @@ class CliTraceCallbacks(TraceCallbacks):
         self.stack.close()
         console.print(f"💾 Trace saved to [cyan]{self.output_path}[/cyan]")
 
+
 @app.command(help="Trace a model.")
 def trace(
     model_path: Annotated[
@@ -557,14 +561,24 @@ def estimate_batchsize(
     ] = None,
 ) -> None:
     if vram_gb is not None:
-        mem = vram_gb * 1024 * 1024 * 1024
-    elif (mem := get_default_device_memory()) is None:
+        # note that in practice GPUs use GiB in their docs, e.g. H100 actually has 85GB of memory
+        mem_bytes = vram_gb * 1000 * 1000 * 1000
+    elif (mem_bytes := get_default_device_bytes()) is None:
         err_console.print("Cannot get the default device's memory stats, use --vram-gb")
         raise Exit(1)
 
+    usable_mem = get_usable_memory_from_bytes(mem_bytes)
+
     callbacks_type = CliEstimateBatchsizeCallbacks
 
-    _estimate_batchsize(model_path, mem, max_input_length, max_output_length, num_logits_per_token, callbacks_type)
+    _estimate_batchsize(
+        model_path,
+        usable_mem,
+        max_input_length,
+        max_output_length,
+        num_logits_per_token,
+        callbacks_type,
+    )
 
 
 @dataclass

{lalamo-0.6.2 → lalamo-0.6.4}/lalamo/speculator/estimator.py

@@ -1,5 +1,6 @@
 import functools
 import itertools
+import os
 from collections.abc import Callable
 from typing import NamedTuple
 
@@ -9,11 +10,38 @@ import jax.numpy as jnp
 from lalamo.models import LanguageModel
 
 
-def get_default_device_memory() -> int | None:
+def get_default_device_bytes() -> 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"]
+
+    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
+
+    # 500mb is seemingly the usually observed overhead; this tries to match the actual capacity of the gpu
+    # so it should correspond to something you'd see in nvidia-smi
+    memory_limit = memory_stats["bytes_limit"] / min(mem_fraction, 1.0) + (500 * 1000 * 1000)
+
+    return get_usable_memory_from_bytes(memory_limit)
+
+
+def get_usable_memory_from_bytes(limit_bytes: int) -> int:
+    # JAX allocates a bit more than it needs, so we discount it by some safety factor
+    return int(limit_bytes * 0.95)
 
 
 def estimate_memory_from_batchsize(
@@ -30,14 +58,14 @@ 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-0.6.2 → lalamo-0.6.4}/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 → lalamo-0.6.4}/lalamo.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lalamo
-Version: 0.6.2
+Version: 0.6.4
 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