lalamo 0.6.3__tar.gz → 0.6.5__tar.gz

{lalamo-0.6.3 → lalamo-0.6.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lalamo
-Version: 0.6.3
+Version: 0.6.5
 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-0.6.3 → lalamo-0.6.5}/lalamo/__init__.py

@@ -32,7 +32,7 @@ from lalamo.speculator import (
     SpeculatorTrainingEvent,
 )
 
-__version__ = "0.6.3"
+__version__ = "0.6.5"
 
 __all__ = [
     "AssistantMessage",

{lalamo-0.6.3 → lalamo-0.6.5}/lalamo/common.py

@@ -1,9 +1,11 @@
+import warnings
 from collections import defaultdict
-from collections.abc import Mapping, Sequence
+from collections.abc import Callable, Iterable, Mapping, Sequence
 from typing import cast
 
 import jax.numpy as jnp
 from jax._src.api import ShapeDtypeStruct
+from jax.errors import JaxRuntimeError
 from jaxtyping import Array, DTypeLike
 
 from lalamo.utils import MapDictValues, MapSequence
@@ -11,8 +13,10 @@ from lalamo.utils import MapDictValues, MapSequence
 __all__ = [
     "DEFAULT_PRECISION",
     "ArrayLike",
+    "LalamoWarning",
     "ParameterPath",
     "ParameterTree",
+    "decrease_batchsize_on_oom",
     "dummy_array",
     "flatten_parameters",
     "require_array",
@@ -23,6 +27,10 @@ __all__ = [
 DEFAULT_PRECISION: DTypeLike = jnp.bfloat16
 
 
+class LalamoWarning(UserWarning):
+    """Custom warning class for Lalamo-specific warnings."""
+
+
 type ArrayLike = Array | ShapeDtypeStruct
 
 
@@ -121,3 +129,49 @@ class ParameterPath(str):
         if not self:
             return ParameterPath(str(other))
         return ParameterPath(self + "." + str(other))
+
+
+def decrease_batchsize_on_oom[T](
+    fn: Callable[[int], Iterable[T]],
+    starting_batch_size: int,
+) -> Iterable[T]:
+    """
+    Execute fn(batch_size) with automatic batch size reduction on OOM.
+    Only reduces batch size if OOM happened on the first batch.
+
+    Args:
+        fn: Function that takes batch_size and returns an iterable
+        starting_batch_size: Initial batch size to try
+
+    Yields:
+        Results from fn(batch_size)
+
+    Raises:
+        JaxRuntimeError: If OOM occurs after first batch completes or at batch_size=1
+    """
+    first_batch_completed = False
+    effective_batch_size = starting_batch_size
+
+    while True:
+        try:
+            for result in fn(effective_batch_size):
+                yield result
+
+                # as soon as we yielded we are not allowed to retry anymore
+                # to make sure we don't ever miss/duplicate outputs
+                first_batch_completed = True
+            break
+        except JaxRuntimeError:
+            if first_batch_completed:
+                raise
+            # because OOM's sometimes generate stuff that won't be garbage collected,
+            # we need to be very aggressive with decreasing batchsize here
+            new_bs = max(int(0.7 * effective_batch_size - 1), 1)
+            if new_bs == 1 and effective_batch_size == 1:
+                raise
+            warnings.warn(
+                f"OOM detected. Reducing batch size {effective_batch_size} -> {new_bs}.",
+                LalamoWarning,
+                stacklevel=3,
+            )
+            effective_batch_size = new_bs

{lalamo-0.6.3 → lalamo-0.6.5}/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-0.6.3 → lalamo-0.6.5}/lalamo/speculator/estimator.py

@@ -10,7 +10,7 @@ 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", "")
@@ -26,20 +26,22 @@ def get_default_device_memory() -> int | None:
     if memory_stats is None or "bytes_limit" not in memory_stats:
         return None
 
-    # 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
+    # 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 * 0.93)
 
 
 def estimate_memory_from_batchsize(

lalamo-0.6.5/lalamo/speculator/inference.py

@@ -0,0 +1,112 @@
+import functools
+from collections.abc import Callable, Iterable
+from itertools import batched, chain, islice
+from typing import NamedTuple
+
+import jax
+import jax.numpy as jnp
+from jax._src.stages import Compiled
+
+from lalamo.common import decrease_batchsize_on_oom
+from lalamo.data.lalamo_completions import LalamoCompletion
+from lalamo.data.utils import get_prefixes_ending_in_user_message
+from lalamo.message_processor import Message
+from lalamo.models import LanguageModel
+
+
+class CollectTracesEvent(NamedTuple):
+    sequences_processed: int
+    tokens_generated: int
+
+
+def inference_collect_traces(
+    model: LanguageModel,
+    conversations: Iterable[Iterable[Message]],
+    num_top_logits_to_collect: int = 8,
+    batch_size: int = 1,
+    max_input_length: int = 1024,
+    max_output_length: int = 1024,
+    tokens_to_generate: int | None = None,
+    progress_callback: Callable[[CollectTracesEvent], None] | None = None,
+) -> Iterable[LalamoCompletion]:
+    def make_generate_tokens_compiled(batch_size: int) -> Compiled:
+        return (
+            jax.jit(
+                functools.partial(
+                    LanguageModel.generate_tokens,
+                    max_output_length=max_output_length,
+                    num_top_logits_to_return=num_top_logits_to_collect,
+                ),
+            )
+            .lower(
+                model,
+                prompt_token_ids=jax.ShapeDtypeStruct((batch_size, max_input_length), jnp.int32),
+                prompt_lengths_without_padding=jax.ShapeDtypeStruct((batch_size,), jnp.int32),
+            )
+            # 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": "0"})
+        )
+
+    prefixes = chain.from_iterable(map(get_prefixes_ending_in_user_message, conversations))
+
+    tokenized_prefixes = map(model.message_processor.tokenize_request, prefixes)
+    filtered_prefixes = filter(lambda conv: len(conv) <= max_input_length, tokenized_prefixes)
+
+    test_batch = list(islice(filtered_prefixes, batch_size))
+
+    def collect_traces_body(batch_size: int) -> Iterable[LalamoCompletion]:
+        tokens_generated, sequences_processed = 0, 0
+        generate_tokens_compiled = make_generate_tokens_compiled(batch_size)
+        for real_batch in batched(chain(test_batch, filtered_prefixes), n=batch_size):
+            batch_padding = batch_size - len(real_batch)
+            batch = (*real_batch, *(([0],) * batch_padding))
+
+            length_without_padding = jnp.array(list(map(len, batch)))
+
+            padded = jnp.array(
+                [
+                    jnp.pad(jnp.array(tokens), (0, max_input_length - len(tokens)), constant_values=0)
+                    for tokens in batch
+                ],
+            )
+
+            generated = generate_tokens_compiled(
+                model,
+                prompt_token_ids=padded,
+                prompt_lengths_without_padding=length_without_padding,
+            )
+
+            assert generated.top_k_token_ids is not None and generated.top_k_token_logits is not None
+
+            for conv_idx in range(len(real_batch)):
+                token_ids = generated.token_ids[conv_idx].tolist()
+                seqlen = next((i + 1 for i, t in enumerate(token_ids) if t in model.stop_token_ids), len(token_ids))
+                if tokens_to_generate is not None:
+                    seqlen = min(seqlen, tokens_to_generate - tokens_generated)
+                tokens_generated += seqlen
+                sequences_processed += 1
+
+                token_ids = token_ids[:seqlen]
+                token_logits_ids = generated.top_k_token_ids[conv_idx, : len(token_ids)].tolist()
+                token_logits_values = generated.top_k_token_logits[conv_idx, : len(token_ids)].tolist()
+                token_logits = [
+                    dict(zip(keys, values, strict=True))
+                    for keys, values in zip(token_logits_ids, token_logits_values, strict=True)
+                ]
+
+                yield LalamoCompletion(batch[conv_idx], token_ids, token_logits)
+
+                if tokens_to_generate is not None and tokens_generated >= tokens_to_generate:
+                    break
+
+            if progress_callback is not None:
+                progress_callback(CollectTracesEvent(sequences_processed, tokens_generated))
+
+            if tokens_to_generate is not None and tokens_generated >= tokens_to_generate:
+                break
+
+    yield from decrease_batchsize_on_oom(collect_traces_body, batch_size)

{lalamo-0.6.3 → lalamo-0.6.5}/lalamo.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lalamo
-Version: 0.6.3
+Version: 0.6.5
 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-0.6.3/lalamo/speculator/inference.py

@@ -1,101 +0,0 @@
-import functools
-from collections.abc import Callable, Iterable
-from itertools import batched, chain
-from typing import NamedTuple
-
-import jax
-import jax.numpy as jnp
-
-from lalamo.data.lalamo_completions import LalamoCompletion
-from lalamo.data.utils import get_prefixes_ending_in_user_message
-from lalamo.message_processor import Message
-from lalamo.models import LanguageModel
-
-
-class CollectTracesEvent(NamedTuple):
-    sequences_processed: int
-    tokens_generated: int
-
-
-def inference_collect_traces(
-    model: LanguageModel,
-    conversations: Iterable[Iterable[Message]],
-    num_top_logits_to_collect: int = 8,
-    batch_size: int = 1,
-    max_input_length: int = 1024,
-    max_output_length: int = 1024,
-    tokens_to_generate: int | None = None,
-    progress_callback: Callable[[CollectTracesEvent], None] | None = None,
-) -> Iterable[LalamoCompletion]:
-    generate_tokens_compiled = (
-        jax.jit(
-            functools.partial(
-                LanguageModel.generate_tokens,
-                max_output_length=max_output_length,
-                num_top_logits_to_return=num_top_logits_to_collect,
-            ),
-        )
-        .lower(
-            model,
-            prompt_token_ids=jax.ShapeDtypeStruct((batch_size, max_input_length), jnp.int32),
-            prompt_lengths_without_padding=jax.ShapeDtypeStruct((batch_size,), jnp.int32),
-        )
-        # 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))
-
-    tokenized_prefixes = map(model.message_processor.tokenize_request, prefixes)
-    filtered_prefixes = filter(lambda conv: len(conv) <= max_input_length, tokenized_prefixes)
-
-    tokens_generated, sequences_processed = 0, 0
-
-    for real_batch in batched(filtered_prefixes, n=batch_size):
-        batch_padding = batch_size - len(real_batch)
-        batch = (*real_batch, *(([0],) * batch_padding))
-
-        length_without_padding = jnp.array(list(map(len, batch)))
-
-        padded = jnp.array(
-            [jnp.pad(jnp.array(tokens), (0, max_input_length - len(tokens)), constant_values=0) for tokens in batch],
-        )
-
-        generated = generate_tokens_compiled(
-            model,
-            prompt_token_ids=padded,
-            prompt_lengths_without_padding=length_without_padding,
-        )
-
-        assert generated.top_k_token_ids is not None and generated.top_k_token_logits is not None
-
-        for conv_idx in range(len(real_batch)):
-            token_ids = generated.token_ids[conv_idx].tolist()
-            seqlen = next((i + 1 for i, t in enumerate(token_ids) if t in model.stop_token_ids), len(token_ids))
-            if tokens_to_generate is not None:
-                seqlen = min(seqlen, tokens_to_generate - tokens_generated)
-            tokens_generated += seqlen
-            sequences_processed += 1
-
-            token_ids = token_ids[:seqlen]
-            token_logits_ids = generated.top_k_token_ids[conv_idx, : len(token_ids)].tolist()
-            token_logits_values = generated.top_k_token_logits[conv_idx, : len(token_ids)].tolist()
-            token_logits = [
-                dict(zip(keys, values, strict=True))
-                for keys, values in zip(token_logits_ids, token_logits_values, strict=True)
-            ]
-
-            yield LalamoCompletion(batch[conv_idx], token_ids, token_logits)
-
-            if tokens_to_generate is not None and tokens_generated >= tokens_to_generate:
-                break
-
-        if progress_callback is not None:
-            progress_callback(CollectTracesEvent(sequences_processed, tokens_generated))
-
-        if tokens_to_generate is not None and tokens_generated >= tokens_to_generate:
-            break