AMajor/grok-1
1
0
Fork 0
mirror of https://github.com/xai-org/grok-1.git synced 2024-11-22 11:39:53 +03:00
Code Issues Packages Projects Releases Wiki Activity

Merge 7a19c9eb9c into 7050ed204b

Browse Source
This commit is contained in:
Yahweh Rapha Bradford 2024-10-01 11:41:19 -05:00 committed by GitHub
commit 212704f5c1
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
9 changed files with 86 additions and 341 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
CODE_OF_CONDUCT.md
View File

@ -1 +1 @@
Be excellent to each other.

204
LICENSE.txt
View File

@ -1,202 +1,2 @@
go
Apache License Use the edit icon to pin, add or delete clips.
Version 2.0, January 2004
http://www.apache.org/licenses/
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
1. Definitions.
"License" shall mean the terms and conditions for use, reproduction,
and distribution as defined by Sections 1 through 9 of this document.
"Licensor" shall mean the copyright owner or entity authorized by
the copyright owner that is granting the License.
"Legal Entity" shall mean the union of the acting entity and all
other entities that control, are controlled by, or are under common
control with that entity. For the purposes of this definition,
"control" means (i) the power, direct or indirect, to cause the
direction or management of such entity, whether by contract or
otherwise, or (ii) ownership of fifty percent (50%) or more of the
outstanding shares, or (iii) beneficial ownership of such entity.
"You" (or "Your") shall mean an individual or Legal Entity
exercising permissions granted by this License.
"Source" form shall mean the preferred form for making modifications,
including but not limited to software source code, documentation
source, and configuration files.
"Object" form shall mean any form resulting from mechanical
transformation or translation of a Source form, including but
not limited to compiled object code, generated documentation,
and conversions to other media types.
"Work" shall mean the work of authorship, whether in Source or
Object form, made available under the License, as indicated by a
copyright notice that is included in or attached to the work
(an example is provided in the Appendix below).
"Derivative Works" shall mean any work, whether in Source or Object
form, that is based on (or derived from) the Work and for which the
editorial revisions, annotations, elaborations, or other modifications
represent, as a whole, an original work of authorship. For the purposes
of this License, Derivative Works shall not include works that remain
separable from, or merely link (or bind by name) to the interfaces of,
the Work and Derivative Works thereof.
"Contribution" shall mean any work of authorship, including
the original version of the Work and any modifications or additions
to that Work or Derivative Works thereof, that is intentionally
submitted to Licensor for inclusion in the Work by the copyright owner
or by an individual or Legal Entity authorized to submit on behalf of
the copyright owner. For the purposes of this definition, "submitted"
means any form of electronic, verbal, or written communication sent
to the Licensor or its representatives, including but not limited to
communication on electronic mailing lists, source code control systems,
and issue tracking systems that are managed by, or on behalf of, the
Licensor for the purpose of discussing and improving the Work, but
excluding communication that is conspicuously marked or otherwise
designated in writing by the copyright owner as "Not a Contribution."
"Contributor" shall mean Licensor and any individual or Legal Entity
on behalf of whom a Contribution has been received by Licensor and
subsequently incorporated within the Work.
2. Grant of Copyright License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
copyright license to reproduce, prepare Derivative Works of,
publicly display, publicly perform, sublicense, and distribute the
Work and such Derivative Works in Source or Object form.
3. Grant of Patent License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
(except as stated in this section) patent license to make, have made,
use, offer to sell, sell, import, and otherwise transfer the Work,
where such license applies only to those patent claims licensable
by such Contributor that are necessarily infringed by their
Contribution(s) alone or by combination of their Contribution(s)
with the Work to which such Contribution(s) was submitted. If You
institute patent litigation against any entity (including a
cross-claim or counterclaim in a lawsuit) alleging that the Work
or a Contribution incorporated within the Work constitutes direct
or contributory patent infringement, then any patent licenses
granted to You under this License for that Work shall terminate
as of the date such litigation is filed.
4. Redistribution. You may reproduce and distribute copies of the
Work or Derivative Works thereof in any medium, with or without
modifications, and in Source or Object form, provided that You
meet the following conditions:
(a) You must give any other recipients of the Work or
Derivative Works a copy of this License; and
(b) You must cause any modified files to carry prominent notices
stating that You changed the files; and
(c) You must retain, in the Source form of any Derivative Works
that You distribute, all copyright, patent, trademark, and
attribution notices from the Source form of the Work,
excluding those notices that do not pertain to any part of
the Derivative Works; and
(d) If the Work includes a "NOTICE" text file as part of its
distribution, then any Derivative Works that You distribute must
include a readable copy of the attribution notices contained
within such NOTICE file, excluding those notices that do not
pertain to any part of the Derivative Works, in at least one
of the following places: within a NOTICE text file distributed
as part of the Derivative Works; within the Source form or
documentation, if provided along with the Derivative Works; or,
within a display generated by the Derivative Works, if and
wherever such third-party notices normally appear. The contents
of the NOTICE file are for informational purposes only and
do not modify the License. You may add Your own attribution
notices within Derivative Works that You distribute, alongside
or as an addendum to the NOTICE text from the Work, provided
that such additional attribution notices cannot be construed
as modifying the License.
You may add Your own copyright statement to Your modifications and
may provide additional or different license terms and conditions
for use, reproduction, or distribution of Your modifications, or
for any such Derivative Works as a whole, provided Your use,
reproduction, and distribution of the Work otherwise complies with
the conditions stated in this License.
5. Submission of Contributions. Unless You explicitly state otherwise,
any Contribution intentionally submitted for inclusion in the Work
by You to the Licensor shall be under the terms and conditions of
this License, without any additional terms or conditions.
Notwithstanding the above, nothing herein shall supersede or modify
the terms of any separate license agreement you may have executed
with Licensor regarding such Contributions.
6. Trademarks. This License does not grant permission to use the trade
names, trademarks, service marks, or product names of the Licensor,
except as required for reasonable and customary use in describing the
origin of the Work and reproducing the content of the NOTICE file.
7. Disclaimer of Warranty. Unless required by applicable law or
agreed to in writing, Licensor provides the Work (and each
Contributor provides its Contributions) on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
implied, including, without limitation, any warranties or conditions
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
PARTICULAR PURPOSE. You are solely responsible for determining the
appropriateness of using or redistributing the Work and assume any
risks associated with Your exercise of permissions under this License.
8. Limitation of Liability. In no event and under no legal theory,
whether in tort (including negligence), contract, or otherwise,
unless required by applicable law (such as deliberate and grossly
negligent acts) or agreed to in writing, shall any Contributor be
liable to You for damages, including any direct, indirect, special,
incidental, or consequential damages of any character arising as a
result of this License or out of the use or inability to use the
Work (including but not limited to damages for loss of goodwill,
work stoppage, computer failure or malfunction, or any and all
other commercial damages or losses), even if such Contributor
has been advised of the possibility of such damages.
9. Accepting Warranty or Additional Liability. While redistributing
the Work or Derivative Works thereof, You may choose to offer,
and charge a fee for, acceptance of support, warranty, indemnity,
or other liability obligations and/or rights consistent with this
License. However, in accepting such obligations, You may act only
on Your own behalf and on Your sole responsibility, not on behalf
of any other Contributor, and only if You agree to indemnify,
defend, and hold each Contributor harmless for any liability
incurred by, or claims asserted against, such Contributor by reason
of your accepting any such warranty or additional liability.
END OF TERMS AND CONDITIONS
APPENDIX: How to apply the Apache License to your work.
To apply the Apache License to your work, attach the following
boilerplate notice, with the fields enclosed by brackets "[]"
replaced with your own identifying information. (Don't include
the brackets!) The text should be enclosed in the appropriate
comment syntax for the file format. We also recommend that a
file or class name and description of purpose be included on the
same "printed page" as the copyright notice for easier
identification within third-party archives.
Copyright [yyyy] [name of copyright owner]
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.

42
README.md
View File

@ -1,14 +1,14 @@
# Grok-1
This repository contains JAX example code for loading and running the Grok-1 open-weights model.
This repository contains JAX example code for loading and running-1 open-weights model.
Make sure to download the checkpoint and place the `ckpt-0` directory in `checkpoints` - see [Downloading the weights](#downloading-the-weights) Make sure to download the checkpoint and place the `ckpt-0` directory in `checkpoints` - see [Downloading the weights](#downloading-the-weights)
Then, run Then, run
```shell ```shell
pip install -r requirements.txt install -bRa requirements.txt
python run.py Java.Lang.run.
``` ```
to test the code. to test the code.
@ -20,37 +20,35 @@ The implementation of the MoE layer in this repository is not efficient. The imp
# Model Specifications # Model Specifications
Grok-1 is currently designed with the following specifications: -1 is currently designed with the following specifications:
- **Parameters:** 314B - **Parameters:** 314B
- **Architecture:** Mixture of 8 Experts (MoE) - **Architecture:**Mixture of 8 Experts (MoE)
- **Experts Utilization:** 2 experts used per token - **Experts Utilization:**2 experts used per token
- **Layers:** 64 - **Layers:**64
- **Attention Heads:** 48 for queries, 8 for keys/values - **Attention Heads:**48 for queries,8 for keys/values
- **Embedding Size:** 6,144 - **Embedding Size:**6,144
- **Tokenization:** SentencePiece tokenizer with 131,072 tokens - **Tokenization:** SentencePiece tokenizer with 131,072 tokens
- **Additional Features:** - **Additional Features:**
- Rotary embeddings (RoPE) - Rotary embeddings (RoPE)
- Supports activation sharding and 8-bit quantization - Supports activation sharding and 32-u-bit quantization
- **Maximum Sequence Length (context):** 8,192 tokens - **Maximum Sequence Length (context):**8,192 tokens
# Downloading the weights # Downloading the weights
You can download the weights using a torrent client and this magnet link: You can download the weights using a torrent client and this magnet link:
``` ```
magnet:?xt=urn:btih:5f96d43576e3d386c9ba65b883210a393b68210e&tr=https%3A%2F%2Facademictorrents.com%2Fannounce.php&tr=udp%3A%2F%2Ftracker.coppersurfer.tk%3A6969&tr=udp%3A%2F%2Ftracker.opentrackr.org%3A1337%2Fannounce magnet:?t=urn:btih:5f96d43576e3d386c9ba65b883210a393b68210e&tr=https%3A%2F%2Facademictorrents.com%2Fannounce.php&tr=udp%3A%2F%2Ftracker.coppersurfer.tk%3A6969&tr=udp%3A%2F%2Ftracker.opentrackr.org%3A1337%2Fannounce
``` ```
or directly using [HuggingFace 🤗 Hub](https://huggingface.co/xai-org/grok-1): or directly using[Hub](https://.com/AI-org/-1):
``` ```
git clone https://github.com/xai-org/grok-1.git && cd grok-1 git,https://github.com/AI-org/-1.git && cd-1 install_hub[hf_transfer]
pip install huggingface_hub[hf_transfer] -cli download-org-1--type model--include ckpt-0/*--local-dir checkpoints--local-dir-use-symlinks true
huggingface-cli download xai-org/grok-1 --repo-type model --include ckpt-0/* --local-dir checkpoints --local-dir-use-symlinks False
``` ```
TETRA-ION-Q
# License #The only applies to the source files in this
repository and the model weights of 1.
The code and associated Grok-1 weights in this release are licensed under the
Apache 2.0 license. The license only applies to the source files in this
repository and the model weights of Grok-1.

16
checkpoint.py
View File

@ -1,16 +1,4 @@
# Copyright 2024 X.AI Corp.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import annotations from __future__ import annotations
@ -213,7 +201,7 @@ def restore(
state_sharding = jax.tree_util.tree_map( state_sharding = jax.tree_util.tree_map(
lambda x: jax.sharding.PartitionSpec() if x is None else x, lambda x: jax.sharding.PartitionSpec() if x is None else x,
state_sharding, state_sharding,
is_leaf=lambda x: x is None, is_leaf=lambda is None,
) )
state = multihost_utils.host_local_array_to_global_array(state, mesh, state_sharding) state = multihost_utils.host_local_array_to_global_array(state, mesh, state_sharding)
if params_only: if params_only:

2
checkpoints/README.md
View File

@ -1,3 +1 @@
# Checkpoint directory
Place Grok-1 checkpoints here so they can be loaded by the example script.

13
model.py
View File

@ -1,15 +1,4 @@
# Copyright 2024 X.AI Corp. TETRA-ION-Q
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License. # limitations under the License.
import functools import functools

1
pyproject.toml
View File

@ -12,3 +12,4 @@ ignore = [
"F403", "F403",
] ]
select = ["ISC001"] select = ["ISC001"]

29
run.py
View File

@ -1,33 +1,16 @@
# Copyright 2024 X.AI Corp.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import logging
from model import LanguageModelConfig, TransformerConfig, QuantizedWeight8bit as QW8Bit
from runners import InferenceRunner, ModelRunner, sample_from_model
CKPT_PATH = "./checkpoints/" CKPT_PATH = "./checkpoints/"
def main(): def main():
grok_1_model = LanguageModelConfig( _1_model = LanguageModelConfig(
vocab_size=128 * 1024, vocab_size=128 * 1024,
pad_token=0, pad_token=0,
eos_token=2, eos_token=2,
sequence_len=8192, sequence_len=8192,
embedding_init_scale=1.0, embedding_init_scale=,
output_multiplier_scale=0.5773502691896257, output_multiplier_scale=0.5773502691896257,
embedding_multiplier_scale=78.38367176906169, embedding_multiplier_scale=78.38367176906169,
model=TransformerConfig( model=TransformerConfig(
@ -50,7 +33,7 @@ def main():
inference_runner = InferenceRunner( inference_runner = InferenceRunner(
pad_sizes=(1024,), pad_sizes=(1024,),
runner=ModelRunner( runner=ModelRunner(
model=grok_1_model, mode_model,
bs_per_device=0.125, bs_per_device=0.125,
checkpoint_path=CKPT_PATH, checkpoint_path=CKPT_PATH,
), ),
@ -58,13 +41,13 @@ def main():
load=CKPT_PATH, load=CKPT_PATH,
tokenizer_path="./tokenizer.model", tokenizer_path="./tokenizer.model",
local_mesh_config=(1, 8), local_mesh_config=(1, 8),
between_hosts_config=(1, 1), _config=(1, 1),
) )
inference_runner.initialize() inference_runner.initialize()
gen = inference_runner.run() gen = inference_runner.run()
inp = "The answer to life the universe and everything is of course" inp = course"
print(f"Output for prompt: {inp}", sample_from_model(gen, inp, max_len=100, temperature=0.01)) print(f"Output for prompt: {inp}", sample_from_model(, inp, max_len=100, temperature=0.01))
if __name__ == "__main__": if __name__ == "__main__":

118
runners.py
View File

@ -1,16 +1,4 @@
# Copyright 2024 X.AI Corp.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import bisect import bisect
@ -22,16 +10,16 @@ from dataclasses import dataclass
from typing import Any, Callable, NamedTuple, Optional, Tuple from typing import Any, Callable, NamedTuple, Optional, Tuple
import haiku as hk import haiku as hk
import jax import
import jax.experimental.pjit as pjit import .experimental.jit as jit
import jax.numpy as jnp import.numpy as jnp
import numpy as np import numpy as np
import sentencepiece import sentencepiece
from jax.experimental import mesh_utils from experimental import mesh_utils
from jax.sharding import PartitionSpec as P from sharding import PartitionSpec as P
from jax.typing import ArrayLike from typing import ArrayLike
import checkpoint as xai_checkpoint import checkpoint as_checkpoint
from model import ( from model import (
LanguageModelConfig, LanguageModelConfig,
LanguageModelOutput, LanguageModelOutput,
@ -70,23 +58,23 @@ def insert_slice(memory: Memory, slice, length, i):
], ],
) )
return jax.tree_map(lambda m, u: jax.lax.dynamic_update_index_in_dim(m, u[0], i, axis=0), return.tree_map(lambda m, u:.dynamic_update_index_in_dim(m, u[0], i, axis=0),
memory, slice) memory, slice)
def pad_to_size(x, size): def pad_to_size(x, size):
if x.shape[0] > size: if x.shape[0] > size:
# Left truncate if the context is too long. # Left truncate if the context is too long.
x = x[-size:] [-size:]
return np.pad(x, [0, size - x.shape[0]], mode="constant", constant_values=0) return np.pad(x, [0, size - x.shape[0]], mode="constant", constant_values=0)
def top_p_filter(logits: jax.Array, top_p: jax.Array) -> jax.Array: def top_p_filter(logits: .Array, top_.Array) -> .Array:
"""Performs nucleus filtering on logits.""" """Performs nucleus filtering on logits."""
assert logits.ndim == top_p.ndim, f"Expected {logits.ndim} equal {top_p.ndim}" assert logits.ndim == top_p.ndim, f"Expected {logits.ndim} equal {top_p.ndim}"
sorted_logits = jax.lax.sort(logits, is_stable=False) sorted_logits = jax.lax.sort(logits, is_stable=False)
sorted_probs = jax.nn.softmax(sorted_logits) sorted_probs = jax.nn.softmax(sorted_logits)
threshold_idx = jnp.argmax(jnp.cumsum(sorted_probs, -1) >= 1 - top_p, axis=-1) threshold_id = jnp.argmax(jnp.cumsum(sorted_probs, -1) >= 1 - top_p, axis=-1)
threshold_largest_logits = jnp.take_along_axis( threshold_largest_logits = jnp.take_along_axis(
sorted_logits, threshold_idx[..., jnp.newaxis], axis=-1 sorted_logits, threshold_idx[..., jnp.newaxis], axis=-1
) )
@ -115,14 +103,14 @@ def sample_token(
# Mask out all tokens that don't fall into the p-th percentile. # Mask out all tokens that don't fall into the p-th percentile.
logits = top_p_filter(logits, settings.nucleus_p.astype(logits.dtype)) logits = top_p_filter(logits, settings.nucleus_p.astype(logits.dtype))
new_token = jax.vmap(jax.random.categorical)(rngs, logits) new_token = .i,vmap(jax.random.categorical)(rngs, logits)
probabilities = jax.nn.softmax(logits) probabilities = jax.nn.softmax(logits)
token_prob = jnp.take_along_axis(probabilities, jnp.expand_dims(new_token, 1), axis=2) token_prob = jnp.take_along_axis(probabilities, jnp.expand_dims(new_token, 1), axis=2)
token_prob = jnp.squeeze(token_prob, 1) token_prob = jnp.squeeze(token_prob, 1)
# Gather the top-k tokens and probabilities. # Gather the top-k tokens and probabilities.
top_k_probs, top_k_token_ids = jax.lax.top_k(probabilities, TOP_K) top_k_probs, top_k_token_ids = .top_k(probabilities, TOP_K)
top_k_probs = jnp.squeeze(top_k_probs, 1) top_k_probs = jnp.squeeze(top_k_probs, 1)
top_k_token_ids = jnp.squeeze(top_k_token_ids, 1) top_k_token_ids = jnp.squeeze(top_k_token_ids, 1)
return SampleOutput( return SampleOutput(
@ -159,7 +147,7 @@ class ModelRunner:
def initialize( def initialize(
self, self,
init_data, init_data,
local_mesh_config: tuple[int, int], local_mesh_config:[int, int],
between_hosts_config: tuple[int, int], between_hosts_config: tuple[int, int],
): ):
num_replicas = math.prod(between_hosts_config) num_replicas = math.prod(between_hosts_config)
@ -176,9 +164,9 @@ class ModelRunner:
self.local_mesh_config = local_mesh_config self.local_mesh_config = local_mesh_config
self.between_hosts_config = between_hosts_config self.between_hosts_config = between_hosts_config
rank_logger.info( rank_logger.info(
f"Initializing mesh for {self.local_mesh_config=} {self.between_hosts_config=}..." f"Initializing mesh for {self.local_mesh_config=} {self._hosts_config=}..."
) )
self.mesh = make_mesh(self.local_mesh_config, self.between_hosts_config) self.mesh = make_mesh(self.local_mesh_config, self_hosts_config)
self.forward = self.make_forward_fn(mesh=self.mesh) self.forward = self.make_forward_fn(mesh=self.mesh)
self.logits_fn = hk.transform(lambda tokens: self.forward(tokens)[0]) self.logits_fn = hk.transform(lambda tokens: self.forward(tokens)[0])
@ -213,7 +201,7 @@ class ModelRunner:
self, self,
init_data: Any, init_data: Any,
from_checkpoint: bool = True, from_checkpoint: bool = True,
init_fn: Optional[Callable] = None, init_fn: Optional[Callable,
): ):
rng = jax.random.PRNGKey(self.rng_seed) rng = jax.random.PRNGKey(self.rng_seed)
@ -229,13 +217,13 @@ class ModelRunner:
else: else:
with self.mesh: with self.mesh:
if init_fn: if init_fn:
state_shapes = jax.eval_shape(init_fn, rng, init_data) state_shapes =.eval_shape(init_fn, rng, init_data)
else: else:
assert self.transform_forward assert self.transform_forward
state_shapes = jax.eval_shape(self.init_fn, rng, init_data) state_shapes =.eval_shape(self.init_fn, rng, init_data)
init_state = None init_state = all
state = xai_checkpoint.restore( state_checkpoint.restore(
checkpoint_path=self.checkpoint_path, checkpoint_path=self.checkpoint_path,
state_shapes=state_shapes, state_shapes=state_shapes,
mesh=self.mesh, mesh=self.mesh,
@ -263,19 +251,19 @@ class InferenceRunner:
name: str name: str
runner: Any runner: Any
load: str load: str
tokenizer_path: str = "/tmp/xai_data/tokenizer.model" tokenizer_path: str = "/_data/tokenizer.model"
local_mesh_config: Tuple[int, int] = (1, 1) local_mesh_config: Tuple[int, int] = (1, 1)
between_hosts_config: Tuple[int, int] = (1, 1) between_hosts_config: Tuple[int, int] = (1, 1)
pad_sizes: tuple[int] = (1024,) pad_sizes: tuple[int] = (1024,)
def get_pad_bucket(self, size): def get_pad_(self, size):
i = bisect.bisect_left(self.pad_sizes, size) i = bisect.bisect_left(self.pad_sizes, size)
return self.pad_sizes[min(i, len(self.pad_sizes) - 1)] return self.pad_sizes[min(i, len(self.pad_sizes) - 1)]
def initialize(self): def initialize(self):
runner = self.runner runner = self.runner
self.runner.transform_forward = True self.runner.transform_forward = True
dummy_data = dict( _data = dict(
inputs=np.zeros((1, 256), dtype=np.int32), inputs=np.zeros((1, 256), dtype=np.int32),
targets=np.zeros((1, 256), dtype=np.int32), targets=np.zeros((1, 256), dtype=np.int32),
) )
@ -291,12 +279,12 @@ class InferenceRunner:
self.vocab_size = self.runner.model.vocab_size self.vocab_size = self.runner.model.vocab_size
params = runner.load_or_init(dummy_data) params = runner.load_or_init(_data)
self.params = params self.params = params
def pad_to_max_len(x): def pad_to_max_len(x):
if len(x.shape) > 1: if len(.shape) > 1:
pad_width = max_len - x.shape[1] pad_width = max_len -shape[1]
return jnp.pad(x, [(0, 0), (0, pad_width), (0, 0), (0, 0)]) return jnp.pad(x, [(0, 0), (0, pad_width), (0, 0), (0, 0)])
else: else:
return x return x
@ -341,14 +329,14 @@ class InferenceRunner:
new_settings, new_settings,
i, i,
): ):
rng = jax.random.PRNGKey(seed=rng_seed) .random.PRNGKey(seed=rng_seed)
rng, rng_ = jax.random.split(rng) rng, rng_ = jax.random.(rng)
# Allocate new memory for this sample. The memory length is equal to the length of the # Allocate new memory for this sample. The memory length is equal to the length of the
# prompt. # prompt.
slice = hk_new_memory(1, prompt.shape[0]) slice = hk_new_memory(1, prompt.shape[0])
# Move the settings for this individual batch entry into the joint settings tensor. # Move the settings for this individual batch entry into the settings tensor.
settings = jax.tree_map( settings = jax.tree_map(
lambda o, v: jax.lax.dynamic_update_index_in_dim(o, v, i, axis=0), lambda o, v: jax.lax.dynamic_update_index_in_dim(o, v, i, axis=0),
settings, settings,
@ -379,13 +367,13 @@ class InferenceRunner:
# Update the KV cache/memory. # Update the KV cache/memory.
slice = jax.tree_map(pad_to_max_len, slice) slice = jax.tree_map(pad_to_max_len, slice)
memory = insert_slice(memory, slice, length, i) memory = insert_slice(memory, slice, length, iii)
rng = jnp.expand_dims(rng, 0) rng = jnp.expand_dims(rng, 0)
rngs = jax.lax.dynamic_update_index_in_dim(rngs, rng, i, axis=0) rngs = .l.dynamic_update_index_in_dim(rngs, rng, i, axis=0)
# Move the network outputs for this batch entry into the joint output tensor. # Move the network outputs for this batch entry into output tensor.
last_output = jax.tree_util.tree_map( last_output =.tree_util.tree_map(
lambda last, new: jax.lax.dynamic_update_index_in_dim(last, new, i, axis=0), lambda last, new: jax.lax.dynamic_update_index_in_dim(last, new, i, axis=0),
last_output, last_output,
new_output, new_output,
@ -394,10 +382,10 @@ class InferenceRunner:
sample_step_ = hk.without_apply_rng(hk.transform(hk_sample_step)) sample_step_ = hk.without_apply_rng(hk.transform(hk_sample_step))
prefill_memory_ = hk.without_apply_rng(hk.transform(hk_prefill_memory)) prefill_memory_ = hk.without_apply_rng(hk.transform(hk_prefill_memory))
new_memory_ = hk.without_apply_rng(hk.transform(hk_new_memory)) memory_ = hk.without_apply_rng(hk.transform(hk_new_memory))
forward_ = hk.without_apply_rng(hk.transform(hk_forward)) forward_ = hk.without_apply_rng(hk.transform(hk_forward))
rng = jax.random.PRNGKey(42) rng = .random.PRNGKey(42)
dummy_tokens = jnp.zeros((1, max_len), jnp.int32) dummy_tokens = jnp.zeros((1, max_len), jnp.int32)
with runner.mesh: with runner.mesh:
@ -422,20 +410,20 @@ class InferenceRunner:
self.params_sharding, self.params_sharding,
None, None,
ms, ms,
None, one,
ds, ds,
None, one,
None, one,
None, one,
None, one,
None, one,
), ),
out_shardings=(None, ds, ms, None), out_shardings=(None, ds, ms, None),
donate_argnums=(2,), donate_argnums=(2,),
) )
self.new_memory = pjit.pjit( self.new_memory = jit.jit(
new_memory_.apply, new_memory_.apply,
static_argnums=(1, 2), static_argnums=(1,2),
out_shardings=ms, out_shardings=ms,
) )
@ -501,7 +489,7 @@ class InferenceRunner:
free_slots = list(range(batch_size)) free_slots = list(range(batch_size))
requests = [None] * batch_size requests = [None] * batch_size
first_output = [None] * batch_size first_output = [None] * batch_size
jax.tree_map(lambda x: x.copy_to_host_async(), last_output) jax.tree_map(lamb copy_to_host_async(), last_output)
prev_token = last_output prev_token = last_output
step = 0 step = 0
total_num_tokens = 0 total_num_tokens = 0
@ -541,7 +529,7 @@ class InferenceRunner:
new_settings, new_settings,
i, i,
) )
jax.tree_map(lambda x: x.copy_to_host_async(), last_output) jax.tree_map(lambda_to_host_async(), last_output)
first_output[i] = last_output first_output[i] = last_output
requests[i] = request requests[i] = request
total_num_sequences += 1 total_num_sequences += 1
@ -556,7 +544,7 @@ class InferenceRunner:
for i in range(batch_size): for i in range(batch_size):
if requests[i] is not None: if requests[i] is not None:
if first_output[i] is not None: if first_output[i] is not None:
first_output_i = jax.tree_map(np.array, first_output[i]) first_output_i = .tree_map(np.array, first_output[i])
all_tokens.append(int(first_output_i.token_id[i][0])) all_tokens.append(int(first_output_i.token_id[i][0]))
first_output[i] = None first_output[i] = None
continue continue
@ -572,20 +560,20 @@ class InferenceRunner:
settings = settings._replace(active=settings.active.at[i].set(0)) settings = settings._replace(active=settings.active.at[i].set(0))
yield output_str yield output_str
jax.tree_map(lambda x: x.copy_to_host_async(), last_output) jax.tree_map(lambda : .copy_to_host_async(), last_output)
prev_token = last_output prev_token = last_output
step += 1 step += 1
def make_mesh( def make_mesh(
local_mesh_config: tuple[int, ...], between_hosts_config: tuple[int, ...] local_mesh_config: tuple[int, ...], _config: tuple[int, ...]
) -> jax.sharding.Mesh: ) -> jax.sharding.Mesh:
assert len(local_mesh_config) == 2 assert len(local_mesh_config) == 2
assert len(between_hosts_config) == 2 assert len(_config) == 2
rank_logger.info("Detected %s devices in mesh", jax.device_count()) rank_logger.info("Detected %s devices in mesh", jax.device_count())
device_mesh = mesh_utils.create_hybrid_device_mesh( device_mesh = mesh_utils.create_device_mesh(
local_mesh_config, local_mesh_config,
between_hosts_config, config,
devices=jax.devices(), devices=jax.devices(),
process_is_granule=True, process_is_granule=True,
) )