repo_name
stringlengths
1
62
dataset
stringclasses
1 value
lang
stringclasses
11 values
pr_id
int64
1
20.1k
owner
stringlengths
2
34
reviewer
stringlengths
2
39
diff_hunk
stringlengths
15
262k
code_review_comment
stringlengths
1
99.6k
axlearn
github_2023
python
699
apple
markblee
@@ -0,0 +1,334 @@ +# Copyright © 2024 Apple Inc. +"""Utilities for the detailed tool use metrics.""" + +import re +import string +from enum import Enum +from typing import Union + +from typing_extensions import TypeAlias + +Value = Union[str, int, bool, float] +ValueOrListOf: TypeAlias = Union[Value, list[Value]] + +_S...
nit -- a basic for loop seems more readable here? ```suggestion start_index = len(words) for i, word in enumerate(words): if word not in _STOP_WORDS: start_index = i break ```
axlearn
github_2023
python
699
apple
markblee
@@ -0,0 +1,334 @@ +# Copyright © 2024 Apple Inc. +"""Utilities for the detailed tool use metrics.""" + +import re +import string +from enum import Enum +from typing import Union + +from typing_extensions import TypeAlias + +Value = Union[str, int, bool, float] +ValueOrListOf: TypeAlias = Union[Value, list[Value]] + +_S...
Same here.
axlearn
github_2023
python
699
apple
markblee
@@ -0,0 +1,334 @@ +# Copyright © 2024 Apple Inc. +"""Utilities for the detailed tool use metrics.""" + +import re +import string +from enum import Enum +from typing import Union + +from typing_extensions import TypeAlias + +Value = Union[str, int, bool, float] +ValueOrListOf: TypeAlias = Union[Value, list[Value]] + +_S...
```suggestion # Only handling string payloads for lenient evaluation. ``` (A gentle reminder to apply review comments throughout the PR.)
axlearn
github_2023
python
705
apple
markblee
@@ -666,3 +673,96 @@ def _mha_backward( flash_attention.defvjp(_mha_forward, _mha_backward) + + +def check_local_compute_capability(cc):
Does this need to be a public fn? Also annotate types/returns/docstring etc?
axlearn
github_2023
python
705
apple
markblee
@@ -666,3 +673,96 @@ def _mha_backward( flash_attention.defvjp(_mha_forward, _mha_backward) + + +def check_local_compute_capability(cc): + if cuda_versions is None: + raise RuntimeError("cuDNN is not detected.") + for i in range(jax.local_device_count()): + compute_cap = cuda_versions.cuda_comp...
Is this accurate?
axlearn
github_2023
python
705
apple
markblee
@@ -666,3 +673,96 @@ def _mha_backward( flash_attention.defvjp(_mha_forward, _mha_backward) + + +def check_local_compute_capability(cc): + if cuda_versions is None: + raise RuntimeError("cuDNN is not detected.") + for i in range(jax.local_device_count()): + compute_cap = cuda_versions.cuda_comp...
```suggestion # Check if cuDNN is installed. ``` and below
axlearn
github_2023
python
705
apple
markblee
@@ -174,56 +176,59 @@ def ref_fn(q, k, v, bias): chex.assert_trees_all_close(jax_grads, jax_ref_grads, atol=0.05) -# We also include a test for Triton with Pallas, to cross validate the triton -# compatibility with our own implementation. +# We test the cudnn_dot_product_attention against the reference flash_a...
Can we retain the original atol for the float16 case?
axlearn
github_2023
python
705
apple
markblee
@@ -174,56 +176,65 @@ def ref_fn(q, k, v, bias): chex.assert_trees_all_close(jax_grads, jax_ref_grads, atol=0.05) -# We also include a test for Triton with Pallas, to cross validate the triton -# compatibility with our own implementation. +# We test the cudnn_dot_product_attention against the reference flash_a...
We should probably add an `else: raise` or similar. If someone parameterizes a different dtype, we don't want to inadvertently skip the check.
axlearn
github_2023
python
705
apple
markblee
@@ -174,56 +176,65 @@ def ref_fn(q, k, v, bias): chex.assert_trees_all_close(jax_grads, jax_ref_grads, atol=0.05) -# We also include a test for Triton with Pallas, to cross validate the triton -# compatibility with our own implementation. +# We test the cudnn_dot_product_attention against the reference flash_a...
Same here.
axlearn
github_2023
python
701
apple
markblee
@@ -211,12 +212,21 @@ def __init__( utils.validate_float_dtype(cfg.train_dtype) # Create the device mesh. - self._step_log( - "Devices: global=%s local=%s %s", - jax.device_count(), - jax.local_device_count(), - [device.platform for device in ja...
nit -- you could also just assign `devices,local_devices` and log after the conditional.
axlearn
github_2023
python
701
apple
markblee
@@ -49,7 +52,7 @@ def _compile_and_dump_programs( compile_topology: Optional[str], compile_topology_num_slices: int = 1, ): - with set_data_dir("FAKE"): + with set_data_dir(FLAGS.data_dir):
nit -- We could also read from env rather than taking a new flag.
axlearn
github_2023
python
696
apple
markblee
@@ -1059,19 +1059,25 @@ def _maybe_stop_or_start_tracing( def select_mesh_config(trainer_config: SpmdTrainer.Config, *, mesh_selector: str): - """Selects a mesh rule (if one matches `mesh_selector` to override mesh config. + """Selects a mesh rule (if one matches mesh_selector to override mesh config. - ...
Can we retain the backticks for code formatting?
axlearn
github_2023
python
696
apple
markblee
@@ -0,0 +1,151 @@ +"""Defines trainer config modifiers, which will be used in model definitions."""
Missing copyright.
axlearn
github_2023
python
696
apple
markblee
@@ -0,0 +1,151 @@ +"""Defines trainer config modifiers, which will be used in model definitions.""" + +from typing import Dict, List, Optional, Union + +from axlearn.common import config +from axlearn.common.base_layer import RematSpec +from axlearn.common.config import REQUIRED, ConfigModifier, ConfigOr, Required, con...
```suggestion cfg: The trainer config to be modified. ``` The types are already specified in the signature.
axlearn
github_2023
python
696
apple
markblee
@@ -0,0 +1,151 @@ +"""Defines trainer config modifiers, which will be used in model definitions.""" + +from typing import Dict, List, Optional, Union + +from axlearn.common import config +from axlearn.common.base_layer import RematSpec +from axlearn.common.config import REQUIRED, ConfigModifier, ConfigOr, Required, con...
```suggestion The modified trainer config. ```
axlearn
github_2023
python
696
apple
markblee
@@ -0,0 +1,151 @@ +"""Defines trainer config modifiers, which will be used in model definitions.""" + +from typing import Dict, List, Optional, Union + +from axlearn.common import config +from axlearn.common.base_layer import RematSpec +from axlearn.common.config import REQUIRED, ConfigModifier, ConfigOr, Required, con...
Do we need to support `None`? The caller can just avoid applying the modifier if they don't want to enable it.
axlearn
github_2023
python
696
apple
markblee
@@ -0,0 +1,151 @@ +"""Defines trainer config modifiers, which will be used in model definitions.""" + +from typing import Dict, List, Optional, Union + +from axlearn.common import config +from axlearn.common.base_layer import RematSpec +from axlearn.common.config import REQUIRED, ConfigModifier, ConfigOr, Required, con...
A suggestion on naming: ```suggestion class RematSpecModifier(ConfigModifier): ``` It may be more consistent/discoverable to name them as `XModifier`.
axlearn
github_2023
python
696
apple
markblee
@@ -0,0 +1,150 @@ +# Copyright © 2023 Apple Inc. + +"""Defines trainer config modifiers, which will be used in model definitions.""" + +from typing import Dict, List, Optional, Union + +from axlearn.common import config +from axlearn.common.base_layer import RematSpec +from axlearn.common.config import REQUIRED, Config...
Could we apply the same comment re. None here? Also, should we spell out how users should configure these, e.g. with a config docstring: ``` @config_class class Config(ConfigModifier.Config): """Configures RematSpecModifier. Attributes: remat_policies: A mapping fr...
axlearn
github_2023
python
696
apple
markblee
@@ -0,0 +1,150 @@ +# Copyright © 2023 Apple Inc. + +"""Defines trainer config modifiers, which will be used in model definitions.""" + +from typing import Dict, List, Optional, Union + +from axlearn.common import config +from axlearn.common.base_layer import RematSpec +from axlearn.common.config import REQUIRED, Config...
Likewise with the comments on typing/casing? ```suggestion cfg: The trainer config to be modified. ```
axlearn
github_2023
python
696
apple
markblee
@@ -0,0 +1,150 @@ +# Copyright © 2023 Apple Inc. + +"""Defines trainer config modifiers, which will be used in model definitions.""" + +from typing import Dict, List, Optional, Union + +from axlearn.common import config +from axlearn.common.base_layer import RematSpec +from axlearn.common.config import REQUIRED, Config...
```suggestion class GradientAccumulationModifier(ConfigModifier): ```
axlearn
github_2023
python
696
apple
markblee
@@ -102,6 +105,26 @@ def sharding(self) -> jax.sharding.Sharding: NestedTensorSpec = Optional[Union[TensorSpec, dict[str, Any]]] +def offload_dots_saveble(offload_src, offload_dst):
Missing types/returns? ```suggestion def offload_dots_saveable(offload_src, offload_dst): ```
axlearn
github_2023
python
696
apple
markblee
@@ -102,6 +105,26 @@ def sharding(self) -> jax.sharding.Sharding: NestedTensorSpec = Optional[Union[TensorSpec, dict[str, Any]]] +def offload_dots_saveble(offload_src, offload_dst): + """Extract and combine the policy from save_and_offload_only_these_names and dots_saveable. + https://github.com/google/jax/b...
```suggestion """Extract and combine the policy from save_and_offload_only_these_names and dots_saveable. https://github.com/google/jax/blob/e3110c18f8bce83901cff42458d4204df9e3abeb/jax/_src/ad_checkpoint.py#L151 This would remove the need to match the names for activation tensors. Args:...
axlearn
github_2023
python
696
apple
markblee
@@ -102,6 +105,26 @@ def sharding(self) -> jax.sharding.Sharding: NestedTensorSpec = Optional[Union[TensorSpec, dict[str, Any]]] +def offload_dots_saveble(offload_src, offload_dst): + """Extract and combine the policy from save_and_offload_only_these_names and dots_saveable. + https://github.com/google/jax/b...
```suggestion offload_src: The source device for offloading. offload_dst: The target device for offloading. ```
axlearn
github_2023
python
696
apple
markblee
@@ -102,6 +105,26 @@ def sharding(self) -> jax.sharding.Sharding: NestedTensorSpec = Optional[Union[TensorSpec, dict[str, Any]]] +def offload_dots_saveble(offload_src, offload_dst): + """Extract and combine the policy from save_and_offload_only_these_names and dots_saveable. + https://github.com/google/jax/b...
Do we need this?
axlearn
github_2023
python
696
apple
markblee
@@ -0,0 +1,150 @@ +# Copyright © 2023 Apple Inc. + +"""Defines trainer config modifiers, which will be used in model definitions.""" + +from typing import Dict, List, Optional, Union + +from axlearn.common import config +from axlearn.common.base_layer import RematSpec +from axlearn.common.config import REQUIRED, Config...
```suggestion ```
axlearn
github_2023
python
696
apple
markblee
@@ -0,0 +1,150 @@ +# Copyright © 2023 Apple Inc. + +"""Defines trainer config modifiers, which will be used in model definitions.""" + +from typing import Dict, List, Optional, Union + +from axlearn.common import config +from axlearn.common.base_layer import RematSpec +from axlearn.common.config import REQUIRED, Config...
Same comment as above.
axlearn
github_2023
python
696
apple
markblee
@@ -0,0 +1,150 @@ +# Copyright © 2023 Apple Inc. + +"""Defines trainer config modifiers, which will be used in model definitions.""" + +from typing import Dict, List, Optional, Union + +from axlearn.common import config +from axlearn.common.base_layer import RematSpec +from axlearn.common.config import REQUIRED, Config...
Same comment as above.
axlearn
github_2023
python
696
apple
markblee
@@ -0,0 +1,150 @@ +# Copyright © 2023 Apple Inc. + +"""Defines trainer config modifiers, which will be used in model definitions.""" + +from typing import Dict, List, Optional, Union + +from axlearn.common import config +from axlearn.common.base_layer import RematSpec +from axlearn.common.config import REQUIRED, Config...
Also here.
axlearn
github_2023
python
696
apple
markblee
@@ -0,0 +1,150 @@ +# Copyright © 2023 Apple Inc. + +"""Defines trainer config modifiers, which will be used in model definitions.""" + +from typing import Dict, List, Optional, Union + +from axlearn.common import config +from axlearn.common.base_layer import RematSpec +from axlearn.common.config import REQUIRED, Config...
Since we specify the config modifier as `ConfigOr[ConfigModifier]`, we should use ```suggestion maybe_instantiate(cfg_modifier) for cfg_modifier in cfg.config_modifiers ```
axlearn
github_2023
python
696
apple
markblee
@@ -0,0 +1,150 @@ +# Copyright © 2023 Apple Inc. + +"""Defines trainer config modifiers, which will be used in model definitions.""" + +from typing import Dict, List, Optional, Union + +from axlearn.common import config +from axlearn.common.base_layer import RematSpec +from axlearn.common.config import REQUIRED, Config...
```suggestion config_modifiers: Required[Sequence[ConfigOr[ConfigModifier]]] = REQUIRED ```
axlearn
github_2023
python
696
apple
markblee
@@ -0,0 +1,150 @@ +# Copyright © 2023 Apple Inc. + +"""Defines trainer config modifiers, which will be used in model definitions.""" + +from typing import Dict, List, Optional, Union + +from axlearn.common import config +from axlearn.common.base_layer import RematSpec +from axlearn.common.config import REQUIRED, Config...
```suggestion metric_accumulator: MetricAccumulator.Config = MetricAccumulator.default_config() ```
axlearn
github_2023
python
696
apple
markblee
@@ -0,0 +1,150 @@ +# Copyright © 2023 Apple Inc. + +"""Defines trainer config modifiers, which will be used in model definitions.""" + +from typing import Dict, List, Optional, Union + +from axlearn.common import config +from axlearn.common.base_layer import RematSpec +from axlearn.common.config import REQUIRED, Config...
```suggestion cfg: The trainer config to be modified. ```
axlearn
github_2023
python
696
apple
markblee
@@ -102,6 +105,26 @@ def sharding(self) -> jax.sharding.Sharding: NestedTensorSpec = Optional[Union[TensorSpec, dict[str, Any]]] +def offload_dots_saveble(offload_src, offload_dst): + """Extract and combine the policy from save_and_offload_only_these_names and dots_saveable.
Is this accurate? Seems that `dots_saveable` includes `lax_convolution.conv_general_dilated_p`. Also clarify how it combines `save_and_offload_only_these_names `?
axlearn
github_2023
python
696
apple
markblee
@@ -0,0 +1,180 @@ +# Copyright © 2023 Apple Inc.
```suggestion # Copyright © 2024 Apple Inc. ``` (here and elsewhere)
axlearn
github_2023
python
693
apple
markblee
@@ -659,7 +659,7 @@ class Config(BaseLayer.Config): # Dimension of each attention head. per_head_dim: Required[int] = REQUIRED # Autoregressive cache dtype. Should match the step dtype. - # Needs to match the forward dtype for Repeated layers. If None, infer as config.dtype. + #...
```suggestion # Needs to match the forward dtype for Repeated layers. If None, infer as `self.dtype()`. ```
axlearn
github_2023
python
695
apple
ruomingp
@@ -185,78 +134,169 @@ def test_training_lm_processor_single_example(self, text: str): # The inputs should be one-off the labels. self.assertNestedAllClose(target_labels[:-1], input_ids[1:]) + @parameterized.parameters( + dict( + packing_method=PackingMethodType.EOS_DELIM_MASK, ...
With this change, do we still test the truncation behavior?
axlearn
github_2023
python
695
apple
ruomingp
@@ -142,12 +163,45 @@ def process_batched(inputs: dict[str, Any]) -> dict[str, Any]: target_num_bytes = num_bytes( batched_target_labels, sp_vocab=vocab, newlines_replaced_with=replace_newlines_with ) - return dict( + + result = dict( input_ids=batched_input_ids...
Neat use of tf.scan!
axlearn
github_2023
python
656
apple
samos123
@@ -403,4 +401,24 @@ def make_grad_accum_config( make_grad_accum_config, make_single_host_config_func, 4 ) + if model_size == "70B": + + def make_config_with_act_offload(base_config_name: str) -> SpmdTrainer.Config: + """Make configs for the v5e/v6e tpu w...
Would it call update_model_remat_config twice this way? The first time during initial config of base config and then 2nd time during make_config_with_act_offload
axlearn
github_2023
python
675
apple
ruomingp
@@ -1162,6 +1171,139 @@ def test_only_linear_weight(self): target_state.trainer_state.model["linear2"]["bias"], ) + def _create_fake_state_and_convert(self, scope_mapping: Dict[str, str]): + # Create fake source_state and target_state with nested layers. + source_cfg, source_sta...
Nit: use assertIs(...)?
axlearn
github_2023
python
675
apple
markblee
@@ -791,10 +792,38 @@ def target_to_source(self, target: Builder.State) -> tuple[Builder.State, Any]: return source, aux def source_to_target(self, source: Builder.State, aux: Any) -> Builder.State: - """Load model state from source to target.""" + """Load model state from source to target...
Move this into `_copy_leaf` so we can use separator for the join?
axlearn
github_2023
python
675
apple
markblee
@@ -791,10 +793,41 @@ def target_to_source(self, target: Builder.State) -> tuple[Builder.State, Any]: return source, aux def source_to_target(self, source: Builder.State, aux: Any) -> Builder.State: - """Load model state from source to target.""" + """Load model state from source to target...
```suggestion lambda path, leaf, src_scope=source_scope: _copy_leaf( path, leaf, source_scope=src_scope, ```
axlearn
github_2023
python
674
apple
markblee
@@ -61,8 +61,14 @@ def is_pending(self) -> bool: return self in {CloudBuildStatus.PENDING, CloudBuildStatus.QUEUED, CloudBuildStatus.WORKING} -def get_cloud_build_status(project_id: str, image_name: str) -> Optional[CloudBuildStatus]: - """Get the status of the latest build filter on the image_name. +de...
```suggestion """Gets the status of the latest build by filtering on the build tags or image name. ```
axlearn
github_2023
python
659
apple
markblee
@@ -121,22 +122,47 @@ def filter_for_validation(structure): ) -# pylint: disable-next=redefined-builtin -def save_tf_savables(value_map: Dict[str, Any], *, dir: str): - """Saves TF savables from `value_map` into `dir`.""" +def _upload_dir(src_dir_handle: tempfile.TemporaryDirectory, *, dst_dir: str): + ...
nit -- `makedirs` should work even if it exists.
axlearn
github_2023
python
659
apple
markblee
@@ -272,6 +298,7 @@ def __init__(self, cfg: Config): self._manager = array_serialization.GlobalAsyncCheckpointManager( timeout_secs=cfg.timeout_secs ) + self._executor = futures.ThreadPoolExecutor()
Since we're no longer using a context manager, should we invoke `shutdown` explicitly?
axlearn
github_2023
python
651
apple
cpgaffney1
@@ -922,3 +971,216 @@ def validate_and_restore(*, step: int, ckpt_dir: str): restored_state = state return step, restored_state + + +class OrbaxCheckpointer(BaseCheckpointer): + """A checkpointer that uses orbax CheckpointManager. + + NOTE: While this class uses index files to do additiona...
Use ``` name_format = ocp.path.step.standard_name_format( step_prefix=_STEP_PREFIX, step_format_fixed_length=_STEP_NUM_DIGITS, ) ocp.path.step.build_step_path(directory, name_format, step) ```
axlearn
github_2023
python
651
apple
cpgaffney1
@@ -922,3 +971,216 @@ def validate_and_restore(*, step: int, ckpt_dir: str): restored_state = state return step, restored_state + + +class OrbaxCheckpointer(BaseCheckpointer): + """A checkpointer that uses orbax CheckpointManager. + + NOTE: While this class uses index files to do additiona...
What is the concern with doing this? Orbax doesn't loop in that many additional dependencies.
axlearn
github_2023
python
651
apple
cpgaffney1
@@ -922,3 +971,216 @@ def validate_and_restore(*, step: int, ckpt_dir: str): restored_state = state return step, restored_state + + +class OrbaxCheckpointer(BaseCheckpointer): + """A checkpointer that uses orbax CheckpointManager. + + NOTE: While this class uses index files to do additiona...
My preference would be for doing this check first and raising an error earlier if no steps are found, followed by the try/except. I think it's a bit more readable but up to you.
axlearn
github_2023
python
651
apple
jiya-zhang
@@ -922,3 +953,220 @@ def validate_and_restore(*, step: int, ckpt_dir: str): restored_state = state return step, restored_state + + +class OrbaxCheckpointer(BaseCheckpointer): + """A checkpointer that uses orbax CheckpointManager. + + NOTE: While this class uses index files to do additiona...
Is this method ever called?
axlearn
github_2023
python
651
apple
jiya-zhang
@@ -922,3 +953,220 @@ def validate_and_restore(*, step: int, ckpt_dir: str): restored_state = state return step, restored_state + + +class OrbaxCheckpointer(BaseCheckpointer): + """A checkpointer that uses orbax CheckpointManager. + + NOTE: While this class uses index files to do additiona...
Is this method ever called?
axlearn
github_2023
python
651
apple
jiya-zhang
@@ -922,3 +953,220 @@ def validate_and_restore(*, step: int, ckpt_dir: str): restored_state = state return step, restored_state + + +class OrbaxCheckpointer(BaseCheckpointer): + """A checkpointer that uses orbax CheckpointManager. + + NOTE: While this class uses index files to do additiona...
What are `index` and `tf_ckpt_args`? Does orbax provide similar functionalities, eg. checking tree structure (which I assume is what the index file is for)?
axlearn
github_2023
python
662
apple
markblee
@@ -270,6 +270,7 @@ def sample_decode( Args: prefix: The prefix to use for prompting. Of shape [batch, max_prefix_length]. + The prefix for each example in the batch should begin with the [BOS] token.
nit -- not all vocabs have `[BOS]`, maybe ```suggestion The prefix for each example in the batch should begin with a prompt token (e.g., BOS). ``` ?
axlearn
github_2023
python
662
apple
markblee
@@ -679,6 +679,98 @@ def test_output_logits_modifier(self): decoder = decoder_cfg.instantiate(parent=None) chex.assert_trees_all_close(decoder(5 * jnp.ones(3)), dict(logits=17 * 5 * jnp.ones(3))) + def test_token_scores_match_between_decoded_and_prefix(self): + """Test that token s...
```suggestion This test is intended to detect if the scores from prefill_states do not match up with ```
axlearn
github_2023
python
652
apple
markblee
@@ -172,6 +177,9 @@ def binary_clf_curve( fps = jnp.where(y_pred_diff, fps, jnp.iinfo(jnp.int32).max) fps = jax.lax.cummin(fps, reverse=True) + # Masked entries at the end of thresholds are set to jnp.finfo(jnp.float32).max. + # Reset them to the rightmost unmasked threshold value. + thresholds = j...
Thanks, as discussed offline maybe we should remove this to simplify, since the masked values should not be used anyway, and I'm not sure that `jnp.argsort(weight)[-1]` gives the rightmost unmasked value.
axlearn
github_2023
python
643
apple
ruomingp
@@ -3874,6 +3874,7 @@ def build_remat_spec( ], self_attention: bool = True, feed_forward: bool = False, + offload: bool = False,
Instead of a bool, should we allow the caller to customize `offload_dst` directly? ```suggestion offload_dst: Optional[Literal["pinned_host"]] = None, ``` This will make the API more extensible and closer to the JAX API.
axlearn
github_2023
python
643
apple
ruomingp
@@ -3891,6 +3904,7 @@ def build_remat_spec( stack_cfg: A transformer config. self_attention: Checkpoint self attention layer activations if true. feed_forward: Checkpoint feed-forward layer activations if true. + offload_dst: Destination of remat checkptoing offloading.
Add a link to the JAX documentation on `offset_dst` on the potential values?
axlearn
github_2023
python
643
apple
ruomingp
@@ -188,6 +188,7 @@ def get_trainer_kwargs( num_kv_heads=None if version == Version.V1 else 8, rope_theta=rope_theta, flash_attention=flash_attention, + remat_offload_dst="pinned_host",
Add a comment on the observed MFU and step time?
axlearn
github_2023
python
643
apple
markblee
@@ -3891,6 +3904,10 @@ def build_remat_spec( stack_cfg: A transformer config. self_attention: Checkpoint self attention layer activations if true. feed_forward: Checkpoint feed-forward layer activations if true. + offload_dst: Destination of remat checkptoing offloading. Relevant Maxte...
```suggestion ```
axlearn
github_2023
python
556
apple
ruomingp
@@ -1306,6 +1306,7 @@ def forward( *, key: Optional[Tensor] = None, value: Optional[Tensor] = None, + kv_state: Optional[KVState] = None,
If `kv_state` is not None, how is it used? Should it be passed to `i_proj` below? Otherwise we should throw a NotImplementedError if kv_state is not None.
axlearn
github_2023
python
465
apple
apghml
@@ -94,8 +95,7 @@ def _prune_empty(in_tree: NestedTensor) -> NestedTensor: return prune_tree(in_tree, lambda _, v: isinstance(v, dict) and not v) -@dataclasses.dataclass -class ForwardOutputs: +class ForwardOutputs(NamedTuple):
What is the purpose of this change?
axlearn
github_2023
python
465
apple
apghml
@@ -444,6 +444,153 @@ def _mask_tree(tree: dict, *, keep: dict) -> dict: ) +class MetricsAccumulationOp(NamedTuple):
Axlearn already has metric accumulation classes that are used by evalers. Could those be reused here instead of defining new classes?
axlearn
github_2023
python
465
apple
apghml
@@ -9,7 +9,8 @@ import dataclasses import enum -from typing import Callable, Mapping, Optional, Protocol, Sequence, Tuple +import logging
We typically use absl logging instead of python logging.
axlearn
github_2023
python
465
apple
apghml
@@ -444,6 +444,153 @@ def _mask_tree(tree: dict, *, keep: dict) -> dict: ) +class MetricsAccumulationOp(NamedTuple): + microbatches: int + + def aggregrate(self, x, buffer): + raise NotImplementedError(self) + + def normalize(self, buffer): + raise NotImplementedError(self) + + +class Ar...
Does the existing learner class use this? If not, we should try to be consistent with its API.
axlearn
github_2023
python
465
apple
apghml
@@ -444,6 +444,153 @@ def _mask_tree(tree: dict, *, keep: dict) -> dict: ) +class MetricsAccumulationOp(NamedTuple): + microbatches: int + + def aggregrate(self, x, buffer): + raise NotImplementedError(self) + + def normalize(self, buffer): + raise NotImplementedError(self) + + +class Ar...
I wonder if instead of having a separate learner for microbatching, it would be more flexible to have a generic way of wrapping a ForwardFn so that it uses Jax.lax.map to run the microbatches. Beyond avoiding the need to add a new learner, it would also allow for other microbetching uses outside of learner, eg inferenc...
axlearn
github_2023
python
614
apple
markblee
@@ -0,0 +1,298 @@ +# Copyright © 2024 Apple Inc. +"""This module provides functions to decorate a ForwardFn to allow for a minibatched +version that enables gradient accumulation. +""" +import functools +from typing import Any, Callable, Optional, Tuple + +import jax +import numpy as np +from jax import numpy as jnp + ...
Sorry for missing this earlier, but we should document the other raise conditions too.
axlearn
github_2023
python
544
apple
markblee
@@ -299,6 +299,21 @@ def _start(self): if tier is not None: reserved = str(tier) == "0" logging.info("Found tier=%s in env. Using reserved=%s", tier, reserved) + + # Create labels for vm tier that can be used to group tpu metrics. + # In QRM, vm tier can be one o...
```suggestion if reserved is None: reserved = gcp_settings("reserved_tpu", default=False, required=False) labels = {"bastion_tier": "reserved" if reserved else "spot"} ``` Not sure if `vmtier` label is arbitrary?
axlearn
github_2023
python
544
apple
markblee
@@ -299,6 +299,15 @@ def _start(self): if tier is not None: reserved = str(tier) == "0" logging.info("Found tier=%s in env. Using reserved=%s", tier, reserved) + + # Create labels for vm tier that can be used to group tpu metrics. + # In QRM, vm tier can be one o...
```suggestion # The "reserved" label is used for guaranteed instances and "spot" for other instances (e.g. best-effort or spot instances). ```
axlearn
github_2023
python
544
apple
markblee
@@ -299,6 +299,15 @@ def _start(self): if tier is not None: reserved = str(tier) == "0" logging.info("Found tier=%s in env. Using reserved=%s", tier, reserved) + + # Create labels for vm tier that can be used to group tpu metrics. + # In QRM, vm tier can be one o...
```suggestion # BASTION_TIER env has presendence over the reserved_tpu. ```
axlearn
github_2023
python
592
apple
apghml
@@ -62,6 +62,11 @@ def generate_job_name() -> str: return f"{os.environ['USER'].replace('_', '')}-{uuid.uuid4().hex.lower()[:6]}" +def generate_job_id() -> str: + """Generate a unique job uuid.""" + return str(uuid.uuid4())
IIUC, uuid4 is a random uuid. Do we know if the randomness used is good? Would it make sense to instead use `secrets.token_urlsafe()`, which is cryptographically random?
axlearn
github_2023
others
470
apple
markblee
@@ -75,6 +75,10 @@ ENV RUN_PYTHON_SDK_IN_DEFAULT_ENVIRONMENT=1 RUN pip install .[gcp,dataflow] COPY . . +COPY --from=apache/beam_python3.9_sdk:2.52.0 /opt/apache/beam /opt/apache/beam +ENTRYPOINT ["/opt/apache/beam/boot"]
Do we plan to merge https://github.com/apple/axlearn/pull/384 eventually?
axlearn
github_2023
others
470
apple
markblee
@@ -75,6 +75,10 @@ ENV RUN_PYTHON_SDK_IN_DEFAULT_ENVIRONMENT=1 RUN pip install .[gcp,dataflow] COPY . . +COPY --from=apache/beam_python3.9_sdk:2.52.0 /opt/apache/beam /opt/apache/beam +ENTRYPOINT ["/opt/apache/beam/boot"] + +
nit -- spacing ```suggestion ```
axlearn
github_2023
python
470
apple
markblee
@@ -0,0 +1,205 @@ +"""An Apache Beam example pipeline to run batch inference jobs using a model trained with AXLearn.
```suggestion """An Apache Beam example pipeline to run batch inference jobs using a model trained with AXLearn. ``` Can we also make a separate directory like `axlearn/cloud/gcp/examples` with these scripts, to distinguish from libraries?
axlearn
github_2023
python
470
apple
markblee
@@ -0,0 +1,205 @@ +"""An Apache Beam example pipeline to run batch inference jobs using a model trained with AXLearn. +Command line options: +--module: the same module used for training +--config: the same config used for training +--trainer_dir: location of your checkpoints for inference + +To debug locally: +$ docker...
nit -- Usually, `trainer_dir` refers to the root of this dir. Maybe we can call it `checkpoint_dir`?
axlearn
github_2023
python
470
apple
markblee
@@ -0,0 +1,205 @@ +"""An Apache Beam example pipeline to run batch inference jobs using a model trained with AXLearn. +Command line options: +--module: the same module used for training +--config: the same config used for training +--trainer_dir: location of your checkpoints for inference + +To debug locally: +$ docker...
Comment on why we need this?
axlearn
github_2023
python
470
apple
markblee
@@ -0,0 +1,205 @@ +"""An Apache Beam example pipeline to run batch inference jobs using a model trained with AXLearn. +Command line options: +--module: the same module used for training +--config: the same config used for training +--trainer_dir: location of your checkpoints for inference + +To debug locally: +$ docker...
```suggestion """Defines how to load a custom checkpoint and run inference.""" ```
axlearn
github_2023
python
470
apple
markblee
@@ -0,0 +1,205 @@ +"""An Apache Beam example pipeline to run batch inference jobs using a model trained with AXLearn. +Command line options: +--module: the same module used for training +--config: the same config used for training +--trainer_dir: location of your checkpoints for inference + +To debug locally: +$ docker...
Missing types. Also comment on what `flag_dict` is supposed to represent? I suppose these are the locally-parsed flag values corresponding to trainer flags?
axlearn
github_2023
python
470
apple
markblee
@@ -0,0 +1,205 @@ +"""An Apache Beam example pipeline to run batch inference jobs using a model trained with AXLearn. +Command line options: +--module: the same module used for training +--config: the same config used for training +--trainer_dir: location of your checkpoints for inference + +To debug locally: +$ docker...
Following the [google style guide](https://google.github.io/styleguide/pyguide.html#386-punctuation-spelling-and-grammar), comments should be full sentences with punctuations. (Please also fix elsewhere, thanks!)
axlearn
github_2023
python
470
apple
markblee
@@ -0,0 +1,205 @@ +"""An Apache Beam example pipeline to run batch inference jobs using a model trained with AXLearn. +Command line options: +--module: the same module used for training +--config: the same config used for training +--trainer_dir: location of your checkpoints for inference + +To debug locally: +$ docker...
```suggestion trainer_cfg = trainer_utils.get_trainer_config(flag_values=flag_values) ```
axlearn
github_2023
python
470
apple
markblee
@@ -0,0 +1,205 @@ +"""An Apache Beam example pipeline to run batch inference jobs using a model trained with AXLearn. +Command line options: +--module: the same module used for training +--config: the same config used for training +--trainer_dir: location of your checkpoints for inference + +To debug locally: +$ docker...
```suggestion inference_runner = inference_runner_cfg.instantiate(parent=None) ```
axlearn
github_2023
python
470
apple
markblee
@@ -0,0 +1,205 @@ +"""An Apache Beam example pipeline to run batch inference jobs using a model trained with AXLearn. +Command line options: +--module: the same module used for training +--config: the same config used for training +--trainer_dir: location of your checkpoints for inference + +To debug locally: +$ docker...
nit -- we don't need the extra `method_runner` variable.
axlearn
github_2023
python
470
apple
markblee
@@ -0,0 +1,205 @@ +"""An Apache Beam example pipeline to run batch inference jobs using a model trained with AXLearn. +Command line options: +--module: the same module used for training +--config: the same config used for training +--trainer_dir: location of your checkpoints for inference + +To debug locally: +$ docker...
Should we have a way to configure the PRNGKey?
axlearn
github_2023
python
470
apple
markblee
@@ -0,0 +1,205 @@ +"""An Apache Beam example pipeline to run batch inference jobs using a model trained with AXLearn. +Command line options: +--module: the same module used for training +--config: the same config used for training +--trainer_dir: location of your checkpoints for inference + +To debug locally: +$ docker...
Missing returns?
axlearn
github_2023
python
470
apple
markblee
@@ -0,0 +1,205 @@ +"""An Apache Beam example pipeline to run batch inference jobs using a model trained with AXLearn. +Command line options: +--module: the same module used for training +--config: the same config used for training +--trainer_dir: location of your checkpoints for inference + +To debug locally: +$ docker...
```suggestion Args: batch: A sequence of examples as NestedTensors. model: An instance of a MethodRunner. inference_kwargs: Optional additional keyword arguments for inference. Returns: A list of method runner outputs. ``` In general, we do not...
axlearn
github_2023
python
470
apple
markblee
@@ -0,0 +1,205 @@ +"""An Apache Beam example pipeline to run batch inference jobs using a model trained with AXLearn. +Command line options: +--module: the same module used for training +--config: the same config used for training +--trainer_dir: location of your checkpoints for inference + +To debug locally: +$ docker...
nit -- use fstring for logging.
axlearn
github_2023
python
470
apple
markblee
@@ -0,0 +1,205 @@ +"""An Apache Beam example pipeline to run batch inference jobs using a model trained with AXLearn. +Command line options: +--module: the same module used for training +--config: the same config used for training +--trainer_dir: location of your checkpoints for inference + +To debug locally: +$ docker...
Missing types. Also, comment on what this `PostProcessFn` is supposed to do?
axlearn
github_2023
python
470
apple
markblee
@@ -0,0 +1,205 @@ +"""An Apache Beam example pipeline to run batch inference jobs using a model trained with AXLearn. +Command line options: +--module: the same module used for training +--config: the same config used for training +--trainer_dir: location of your checkpoints for inference + +To debug locally: +$ docker...
nit -- iterate over `fake_input` directly and break (or use `take`) if needed.
axlearn
github_2023
python
470
apple
markblee
@@ -0,0 +1,205 @@ +"""An Apache Beam example pipeline to run batch inference jobs using a model trained with AXLearn. +Command line options: +--module: the same module used for training +--config: the same config used for training +--trainer_dir: location of your checkpoints for inference + +To debug locally: +$ docker...
Seems this is mainly intended for pipeline args? If so, clarify in docstring?
axlearn
github_2023
python
470
apple
markblee
@@ -0,0 +1,105 @@ +"""An Apache Beam example pipeline to run batch inference jobs with a HuggingFace model. +Reference: https://cloud.google.com/dataflow/docs/notebooks/ +run_inference_huggingface#runinference_with_a_pretrained_model_from_hugging_face_hub + +
```suggestion """An Apache Beam example pipeline to run batch inference jobs with a HuggingFace model. Reference: https://cloud.google.com/dataflow/docs/notebooks/ run_inference_huggingface#runinference_with_a_pretrained_model_from_hugging_face_hub ```
axlearn
github_2023
others
470
apple
markblee
@@ -33,6 +33,7 @@ dependencies = [ "tensorstore>=0.1.21", # used for supporting GDA checkpoints "toml", # for config management "typing-extensions==4.9.0", # needed for typing.Protocol. >4.9.0 runs into attribute error `__non_callable_proto_members__`. + "scipy==1.12.0",
I think this should go away with a rebase.
axlearn
github_2023
python
470
apple
markblee
@@ -0,0 +1,202 @@ +"""An Apache Beam example pipeline to run batch inference jobs using a model trained with AXLearn.
I think we require a copyright header on all files.
axlearn
github_2023
python
470
apple
yqwangustc
@@ -0,0 +1,202 @@ +"""An Apache Beam example pipeline to run batch inference jobs using a model trained with AXLearn. + +Command line options: +--module: the same module used for training +--config: the same config used for training +--trainer_dir: location of your checkpoints for inference + +To debug locally: +$ dock...
Looks like this is a new thing. Do you have document pointer for this ModelHandler ?
axlearn
github_2023
others
470
apple
ruomingp
@@ -75,6 +75,9 @@ ENV RUN_PYTHON_SDK_IN_DEFAULT_ENVIRONMENT=1 RUN pip install .[gcp,dataflow] COPY . . +COPY --from=apache/beam_python3.9_sdk:2.52.0 /opt/apache/beam /opt/apache/beam +ENTRYPOINT ["/opt/apache/beam/boot"]
Comment on why we need this step?
axlearn
github_2023
python
470
apple
ruomingp
@@ -0,0 +1,205 @@ +# Copyright © 2024 Google LLC + +"""An Apache Beam example pipeline to run batch inference jobs using a model trained with AXLearn. + +Command line options: +--module: the same module used for training +--config: the same config used for training +--trainer_dir: location of your checkpoints for infer...
Is there a way to use TPU?
axlearn
github_2023
python
470
apple
ruomingp
@@ -0,0 +1,205 @@ +# Copyright © 2024 Google LLC + +"""An Apache Beam example pipeline to run batch inference jobs using a model trained with AXLearn. + +Command line options: +--module: the same module used for training +--config: the same config used for training +--trainer_dir: location of your checkpoints for infer...
Is this method inheriting from the parent class?
axlearn
github_2023
python
224
apple
ruomingp
@@ -130,7 +129,6 @@ def __init__(self, cfg: _BaseLoraAdapter.Config, *, parent: Module): cfg.lora_up.set( input_dim=cfg.rank, output_dim=cfg.output_dim, - param_partition_spec=["model", None],
Is there a test for this change?
axlearn
github_2023
python
578
apple
markblee
@@ -188,6 +187,24 @@ def _aggregate_tool_role_messages(messages: List[Dict[str, Any]]) -> List[Dict[s return aggregated_messages +def _santize_request(request: Dict[str, Any]): + """Santizes request to follow Gemini request rules."""
```suggestion def _format_request(request: Dict[str, Any]): """Formats request to follow Gemini request rules.""" ``` Since it seems more like formatting than sanitization?
axlearn
github_2023
python
578
apple
markblee
@@ -69,20 +65,22 @@ async def async_generate( """ cfg: OpenAIClient.Config = self.config client: AsyncOpenAI = self._client - assert prompt is not None or messages is not None, ValidationError( + prompt = request.get("prompt", None) + messages = request.get("messages", No...
Is this correct? Previously we require one to be non-None, but now we check that one of them is None.
axlearn
github_2023
python
578
apple
markblee
@@ -188,6 +187,24 @@ def _aggregate_tool_role_messages(messages: List[Dict[str, Any]]) -> List[Dict[s return aggregated_messages +def _santize_request(request: Dict[str, Any]): + """Santizes request to follow Gemini request rules.""" + if "messages" in request: + new_messages = [] + for mes...
nit -- could be simpler with list comprehensions.
axlearn
github_2023
python
300
apple
gyin94
@@ -125,26 +125,40 @@ class Foo: def similar_names(name: str, candidates: Iterable[str]) -> List[str]: - """Return a sorted list of candidates that are similar to name.""" - - def overlaps(name: str, key: str) -> float: - """The fraction of 3-char substrings in <name> that appear in key.""" - m...
consider a type hint and return type?
axlearn
github_2023
python
572
apple
patrick-toulme
@@ -305,80 +310,99 @@ def vmap_fn( policy=jax.checkpoint_policies.nothing_saveable, ) def scan_fn( - carry_output_t_1: NestedTensor, - scan_t: Tuple[NestedTensor, NestedTensor, NestedTensor], + carry_in: NestedTensor, + ...
see this PR: https://github.com/markblee/axlearn/pull/1
axlearn
github_2023
python
573
apple
markblee
@@ -272,6 +264,85 @@ async def async_generate_from_requests( del item["async_index"] return responses + @classmethod + def define_flags(cls, fv: flags.FlagValues): + """Defines flags for generator.py.""" + common_kwargs = dict(flag_values=fv, allow_override=True) + ...
Forgot to mention, but for these non-config flags (check_vllm_readiness, debug or repeat_requests_for_n) it may be worth defining them separately in `generator.py`. Feel free to address separately.