AMajor/grok-1
1
0
Fork 0
mirror of https://github.com/xai-org/grok-1.git synced 2024-11-23 20:19:53 +03:00
Code Issues Packages Projects Releases Wiki Activity

Add initial code

Browse Source
This commit is contained in:
Igor Babuschkin 2024-03-14 15:03:58 -07:00
parent 5aabc78af1
commit be76c959fa
11 changed files with 2552 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
CODE_OF_CONDUCT.md Normal file
View File

@ -0,0 +1 @@
Be excellent to each other.

202
LICENSE.txt Normal file
View File

@ -0,0 +1,202 @@
Apache License
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.

33
README.md
View File

@ -1,2 +1,31 @@
# grok-open
Open release of the Grok model
# Grok-1
This repository contains JAX example code for loading and running the Grok-1 open-weights model.
Make sure to download the checkpoint and place `ckpt-0` directory in `checkpoint`.
Then, run
```shell
pip install -r requirements.txt
python run.py
```
to test the code.
The script loads the checkpoint and samples from the model on a test input.
Due to the large size of the model (314B parameters), a machine with enough GPU memory is required to test the model with the example code.
The implementation of the MoE layer in this repository is not efficient. The implementation was chosen to avoid the need for custom kernels to validate the correctness of the model.
# Downloading the weights
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
```
# License
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.

221
checkpoint.py Normal file
View File

@ -0,0 +1,221 @@
# 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
import contextlib
import logging
import math
import os
import pickle
import re
import shutil
import sys
import tempfile
from concurrent.futures import ThreadPoolExecutor, wait
from typing import Any, Optional
import jax
import numpy as np
from jax.experimental import multihost_utils
from model import QuantizedWeight8bit
logger = logging.getLogger(__name__)
rank_logger = logging.getLogger("rank")
# Needed for loading the checkpoint with pickle.
sys.modules['__main__'].QuantizedWeight8bit = QuantizedWeight8bit
@contextlib.contextmanager
def copy_to_shm(file: str):
if file.startswith("/dev/shm/"):
# Nothing to do, the file is already in shared memory.
yield file
return
tmp_dir = "/dev/shm/"
fd, tmp_path = tempfile.mkstemp(dir=tmp_dir)
try:
shutil.copyfile(file, tmp_path)
yield tmp_path
finally:
os.remove(tmp_path)
os.close(fd)
@contextlib.contextmanager
def copy_from_shm(file: str):
tmp_dir = "/dev/shm/"
fd, tmp_path = tempfile.mkstemp(dir=tmp_dir)
try:
yield tmp_path
shutil.copyfile(tmp_path, file)
finally:
os.remove(tmp_path)
os.close(fd)
def fast_unpickle(path: str) -> Any:
with copy_to_shm(path) as tmp_path:
with open(tmp_path, "rb") as f:
return pickle.load(f)
def fast_pickle(obj: Any, path: str) -> None:
with copy_from_shm(path) as tmp_path:
with open(tmp_path, "wb") as f:
pickle.dump(obj, f)
def load_tensors(shaped_arrays, directory, mesh_config, tensor_indices=None):
"""Loads a set of arrays."""
pool = ThreadPoolExecutor(max_workers=32)
fs = list()
num_tensors = 0
num_replicas = 1
data_model_shards = math.prod(mesh_config)
if tensor_indices is None:
iterator = enumerate(shaped_arrays)
else:
iterator = zip(tensor_indices, shaped_arrays)
for i, t in iterator:
if (i % num_replicas) == ((jax.process_index() // data_model_shards) % num_replicas):
idx = (
jax.process_index() // (num_replicas * data_model_shards) * data_model_shards
+ jax.process_index() % data_model_shards
)
fs.append(
pool.submit(fast_unpickle, os.path.join(directory, f"tensor{i:05d}_{idx:03d}"))
)
num_tensors += 1
else:
fs.append(pool.submit(np.zeros, t.shape, dtype=t.dtype))
wait(fs)
return [f.result() for f in fs]
def path_tuple_to_string(path: tuple) -> str:
pieces = []
for elem in path:
if isinstance(elem, jax.tree_util.DictKey):
pieces.append(elem.key)
elif isinstance(elem, jax.tree_util.GetAttrKey):
pieces.append(elem.name)
else:
assert isinstance(elem, (jax.tree_util.FlattenedIndexKey, jax.tree_util.SequenceKey))
return "/".join(pieces)
def get_load_path_str(
init_path_str: str,
load_rename_rules: Optional[list[tuple[str, str]]] = None,
load_exclude_rules: Optional[list[str]] = None,
) -> Optional[str]:
# Exclusion
if load_exclude_rules is not None:
for search_pattern in load_exclude_rules:
if re.search(search_pattern, init_path_str):
return None
# Renaming
load_path_str = init_path_str
if load_rename_rules is not None:
for search_pattern, replacement_pattern in load_rename_rules:
if re.search(search_pattern, load_path_str):
load_path_str = re.sub(search_pattern, replacement_pattern, load_path_str)
break
return load_path_str
def replace_with_load_state(
init_state: Any,
load_state: Any,
load_rename_rules: Optional[list[tuple[str, str]]] = None,
load_exclude_rules: Optional[list[str]] = None,
mesh_config: tuple = (1, 1),
) -> Any:
flatten_load, _ = jax.tree_util.tree_flatten_with_path(load_state)
flatten_init, structure_init = jax.tree_util.tree_flatten_with_path(init_state)
load_map = {path_tuple_to_string(path): tensor for path, tensor in flatten_load}
replaced = []
num_replicas = 1
data_model_shards = math.prod(mesh_config)
for i, (init_path, tensor) in enumerate(flatten_init):
init_path_str = path_tuple_to_string(init_path)
load_path_str = get_load_path_str(init_path_str, load_rename_rules, load_exclude_rules)
if load_path_str is None:
rank_logger.info(f"Excluded from restore: {init_path_str}.")
replaced.append(tensor)
elif load_path_str in load_map:
if load_path_str == init_path_str:
rank_logger.info(f"Restored from ckpt: {init_path_str}.")
else:
rank_logger.info(f"Restored from ckpt: {init_path_str} <-- {load_path_str}.")
replaced.append(load_map[load_path_str])
else:
rank_logger.info(f"Not found in ckpt: {init_path_str}.")
if (i % num_replicas) == ((jax.process_index() // data_model_shards) % num_replicas):
replaced.append(tensor)
else:
replaced.append(np.zeros_like(tensor))
return jax.tree_util.tree_unflatten(structure_init, replaced)
def restore(
checkpoint_path: str,
state_shapes: Any,
mesh,
between_hosts_config,
params_only,
state_sharding,
init_state: Optional[Any] = None,
) -> Any:
ckpt_path = os.path.join(checkpoint_path, "ckpt-0")
rank_logger.info("Loading checkpoint at {}".format(ckpt_path))
ckpt_shapes = state_shapes
ckpt_shapes_with_path, structure = jax.tree_util.tree_flatten_with_path(ckpt_shapes)
ckpt_shapes_flat = [elem[1] for elem in ckpt_shapes_with_path]
loaded_tensors = load_tensors(ckpt_shapes_flat, ckpt_path, between_hosts_config)
state = jax.tree_util.tree_unflatten(structure, loaded_tensors)
# Sanity check to give a better error message.
ckpt_keys = set(state.params.keys())
code_keys = set(state_sharding.params.keys())
if ckpt_keys != code_keys and init_state is None:
missing_in_ckpt = code_keys - ckpt_keys
missing_locally = ckpt_keys - code_keys
raise ValueError(
"Parameters in the code are not matching checkpoint parameters.\n"
"Params missing in checkpoint: {}\nParams missing in code: {}".format(
missing_in_ckpt, missing_locally
)
)
state_sharding = jax.tree_util.tree_map(
lambda x: jax.sharding.PartitionSpec() if x is None else x,
state_sharding,
is_leaf=lambda x: x is None,
)
state = multihost_utils.host_local_array_to_global_array(state, mesh, state_sharding)
if params_only:
state = state.params
return state

3
checkpoints/README.md Normal file
View File

@ -0,0 +1,3 @@
# Checkpoint directory
Place Grok-1 checkpoints here so they can be loaded by the example script.

1398
model.py Normal file
View File

File diff suppressed because it is too large Load Diff

14
pyproject.toml Normal file
View File

@ -0,0 +1,14 @@
[tool.ruff]
indent-width = 4
line-length = 100
[tool.ruff.lint]
ignore = [
"E722",
"E731",
"E741",
"F405",
"E402",
"F403",
]
select = ["ISC001"]

5
requirements.txt Normal file
View File

@ -0,0 +1,5 @@
dm_haiku==0.0.12
-f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html
jax[cuda12_pip]==0.4.25
numpy==1.26.4
sentencepiece==0.2.0

72
run.py Normal file
View File

@ -0,0 +1,72 @@
# 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/"
def main():
grok_1_model = LanguageModelConfig(
vocab_size=128 * 1024,
pad_token=0,
eos_token=2,
sequence_len=8192,
embedding_init_scale=1.0,
output_multiplier_scale=0.5773502691896257,
embedding_multiplier_scale=78.38367176906169,
model=TransformerConfig(
emb_size=48 * 128,
widening_factor=8,
key_size=128,
num_q_heads=48,
num_kv_heads=8,
num_layers=64,
attn_output_multiplier=0.08838834764831845,
shard_activations=True,
# MoE.
num_experts=8,
num_selected_experts=2,
# Activation sharding.
data_axis="data",
model_axis="model",
),
)
inference_runner = InferenceRunner(
pad_sizes=(1024,),
runner=ModelRunner(
model=grok_1_model,
bs_per_device=0.125,
checkpoint_path=CKPT_PATH,
),
name="local",
load=CKPT_PATH,
tokenizer_path="./tokenizer.model",
local_mesh_config=(1, 8),
between_hosts_config=(1, 1),
)
inference_runner.initialize()
gen = inference_runner.run()
inp = "The answer to life the universe and everything is of course"
print(f"Output for prompt: {inp}", sample_from_model(gen, inp, max_len=100, temperature=0.01))
if __name__ == "__main__":
logging.basicConfig(level=logging.INFO)
main()

605
runners.py Normal file
View File

@ -0,0 +1,605 @@
# 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 functools
import logging
import math
import re
from dataclasses import dataclass
from typing import Any, Callable, NamedTuple, Optional, Tuple
import haiku as hk
import jax
import jax.experimental.pjit as pjit
import jax.numpy as jnp
import numpy as np
import sentencepiece
from jax.experimental import mesh_utils
from jax.sharding import PartitionSpec as P
from jax.typing import ArrayLike
import checkpoint as xai_checkpoint
from model import (
LanguageModelConfig,
LanguageModelOutput,
TrainingState,
apply_rules,
Memory,
KVMemory,
)
logger = logging.getLogger(__name__)
rank_logger = logging.getLogger("rank")
TOP_K = 8
class SampleSettings(NamedTuple):
temperature: ArrayLike
nucleus_p: ArrayLike
mask: ArrayLike
# Whether a given batch element is actively used. [B]
active: ArrayLike
class SampleOutput(NamedTuple):
token_id: ArrayLike
prob: ArrayLike
top_k_token_ids: ArrayLike
top_k_probs: ArrayLike
def insert_slice(memory: Memory, slice, length, i):
slice = Memory(
layers=[
KVMemory(layer.k, layer.v, step=jnp.array([length]))
for layer in slice.layers
],
)
return jax.tree_map(lambda m, u: jax.lax.dynamic_update_index_in_dim(m, u[0], i, axis=0),
memory, slice)
def pad_to_size(x, size):
if x.shape[0] > size:
# Left truncate if the context is too long.
x = x[-size:]
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:
"""Performs nucleus filtering on logits."""
assert logits.ndim == top_p.ndim, f"Expected {logits.ndim} equal {top_p.ndim}"
sorted_logits = jax.lax.sort(logits, is_stable=False)
sorted_probs = jax.nn.softmax(sorted_logits)
threshold_idx = jnp.argmax(jnp.cumsum(sorted_probs, -1) >= 1 - top_p, axis=-1)
threshold_largest_logits = jnp.take_along_axis(
sorted_logits, threshold_idx[..., jnp.newaxis], axis=-1
)
assert threshold_largest_logits.shape == logits.shape[:-1] + (1,)
mask = logits >= threshold_largest_logits
# Set unused logits to -inf.
logits = jnp.where(mask, logits, -1e10)
return logits
def sample_token(
rngs: jax.random.PRNGKey,
lm_outputs: LanguageModelOutput,
settings: SampleSettings,
) -> SampleOutput:
# Expand the settings shape to match the logit shape.
settings = SampleSettings(
temperature=jnp.expand_dims(settings.temperature, (1, 2)), # Input [B], output [B, 1, 1].
nucleus_p=jnp.expand_dims(settings.nucleus_p, (1, 2)), # Input [B], output [B, 1, 1].
mask=jnp.expand_dims(settings.mask, 1), # Input [B, V], output [B, 1, V].
active=settings.active, # [B].
)
logits = lm_outputs.logits / settings.temperature.astype(lm_outputs.logits.dtype)
# Mask out all disallowed tokens by assigning them a near-zero probability.
logits = jnp.where(settings.mask, logits, -1e10)
# Mask out all tokens that don't fall into the p-th percentile.
logits = top_p_filter(logits, settings.nucleus_p.astype(logits.dtype))
new_token = jax.vmap(jax.random.categorical)(rngs, logits)
probabilities = jax.nn.softmax(logits)
token_prob = jnp.take_along_axis(probabilities, jnp.expand_dims(new_token, 1), axis=2)
token_prob = jnp.squeeze(token_prob, 1)
# Gather the top-k tokens and probabilities.
top_k_probs, top_k_token_ids = jax.lax.top_k(probabilities, TOP_K)
top_k_probs = jnp.squeeze(top_k_probs, 1)
top_k_token_ids = jnp.squeeze(top_k_token_ids, 1)
return SampleOutput(
new_token,
token_prob,
top_k_token_ids,
top_k_probs,
)
@dataclass
class ModelRunner:
model: LanguageModelConfig
bs_per_device: float = 2.0
load_rename_rules: Optional[list[tuple[str, str]]] = None
load_exclude_rules: Optional[list[str]] = None
rng_seed: int = 42 # Initial rng seed.
transform_forward: bool = False
checkpoint_path: str = ""
def make_forward_fn(self, mesh: Any):
def forward(tokens):
out = self.model.make(mesh=mesh)(tokens)
return out, None
if self.transform_forward:
forward = hk.transform(forward)
return forward
def initialize(
self,
init_data,
local_mesh_config: tuple[int, int],
between_hosts_config: tuple[int, int],
):
num_replicas = math.prod(between_hosts_config)
self.model.initialize()
self.model.fprop_dtype = jnp.bfloat16
num_local_gpus = len(jax.local_devices())
# Calculate the global batch size from the local batch size.
self.batch_size = int(self.bs_per_device * num_local_gpus * num_replicas)
# Calculate the batch size per host from the global batch size.
self.local_batch_size = self.batch_size // jax.process_count()
self.local_mesh_config = local_mesh_config
self.between_hosts_config = between_hosts_config
rank_logger.info(
f"Initializing mesh for {self.local_mesh_config=} {self.between_hosts_config=}..."
)
self.mesh = make_mesh(self.local_mesh_config, self.between_hosts_config)
self.forward = self.make_forward_fn(mesh=self.mesh)
self.logits_fn = hk.transform(lambda tokens: self.forward(tokens)[0])
self.eval_forward = self.make_forward_fn(mesh=self.mesh)
self.logits_eval_fn = hk.transform(lambda tokens: self.eval_forward(tokens)[0])
if self.transform_forward:
self.state_sharding = self.get_state_sharding(init_data)
rank_logger.info(f"State sharding type: {type(self.state_sharding)}")
self.init_fn = pjit.pjit(self.init, out_shardings=self.state_sharding)
def init(self, rng: jax.Array, data) -> TrainingState:
assert self.transform_forward
rng, init_rng = jax.random.split(rng)
params = self.forward.init(init_rng, data["inputs"])
return TrainingState(params=params)
def get_state_sharding(self, init_data):
assert self.transform_forward
rng = jax.random.PRNGKey(self.rng_seed)
rank_logger.info(f"partition rules: {self.model.partition_rules}")
with self.mesh:
shapes = jax.eval_shape(self.init, rng, init_data)
sharding = jax.tree_util.tree_map_with_path(
apply_rules(self.model.partition_rules()),
shapes,
)
return sharding
def load_or_init(
self,
init_data: Any,
from_checkpoint: bool = True,
init_fn: Optional[Callable] = None,
):
rng = jax.random.PRNGKey(self.rng_seed)
if not self.checkpoint_path or not from_checkpoint:
rank_logger.info("Initializing model...")
with self.mesh:
if init_fn is not None:
state = init_fn(rng, init_data)
else:
assert self.transform_forward
state = self.init_fn(rng, init_data)
rank_logger.info("Model state is newly initialized.")
else:
with self.mesh:
if init_fn:
state_shapes = jax.eval_shape(init_fn, rng, init_data)
else:
assert self.transform_forward
state_shapes = jax.eval_shape(self.init_fn, rng, init_data)
init_state = None
state = xai_checkpoint.restore(
checkpoint_path=self.checkpoint_path,
state_shapes=state_shapes,
mesh=self.mesh,
between_hosts_config=self.between_hosts_config,
state_sharding=self.state_sharding,
init_state=init_state,
params_only=True,
)
del init_state
return state
@dataclass
class Request:
prompt: str
temperature: float
nucleus_p: float
rng_seed: int
max_len: int
@dataclass
class InferenceRunner:
name: str
runner: Any
load: str
tokenizer_path: str = "/tmp/xai_data/tokenizer.model"
local_mesh_config: Tuple[int, int] = (1, 1)
between_hosts_config: Tuple[int, int] = (1, 1)
pad_sizes: tuple[int] = (1024,)
def get_pad_bucket(self, size):
i = bisect.bisect_left(self.pad_sizes, size)
return self.pad_sizes[min(i, len(self.pad_sizes) - 1)]
def initialize(self):
runner = self.runner
self.runner.transform_forward = True
dummy_data = dict(
inputs=np.zeros((1, 256), dtype=np.int32),
targets=np.zeros((1, 256), dtype=np.int32),
)
runner.initialize(
dummy_data,
local_mesh_config=self.local_mesh_config,
between_hosts_config=self.between_hosts_config,
)
self.tokenizer = sentencepiece.SentencePieceProcessor(model_file=self.tokenizer_path)
max_len = runner.model.sequence_len
self.vocab_size = self.runner.model.vocab_size
params = runner.load_or_init(dummy_data)
self.params = params
def pad_to_max_len(x):
if len(x.shape) > 1:
pad_width = max_len - x.shape[1]
return jnp.pad(x, [(0, 0), (0, pad_width), (0, 0), (0, 0)])
else:
return x
@functools.lru_cache
def lm():
return runner.model.make(mesh=runner.mesh)
def hk_forward(
tokens,
memory=None,
length=None,
active=None,
) -> LanguageModelOutput:
if memory is not None:
assert active is not None
layers = []
for l in memory.layers:
# Reset steps to 0 for inactive requests to avoid unnecessary computations.
step = jnp.where(active, l.step, jnp.zeros_like(l.step))
layers.append(l._replace(step=step))
memory = memory._replace(layers=layers)
return lm()(tokens, memory, length=length)
def hk_sample_step(rngs, last_output: SampleOutput, memory, settings):
rngs, rngs_ = jax.vmap(jax.random.split, out_axes=1)(rngs)
lm_outputs = hk_forward(last_output.token_id, memory=memory, active=settings.active)
sample_result = sample_token(rngs_, lm_outputs, settings)
return rngs, sample_result, lm_outputs.model_state
def hk_new_memory(batch_size, sequence_len):
return lm().init_memory(batch_size, sequence_len)
def hk_prefill_memory(
rngs,
memory,
settings,
last_output,
prompt,
length,
rng_seed,
new_settings,
i,
):
rng = jax.random.PRNGKey(seed=rng_seed)
rng, rng_ = jax.random.split(rng)
# Allocate new memory for this sample. The memory length is equal to the length of the
# prompt.
slice = hk_new_memory(1, prompt.shape[0])
# Move the settings for this individual batch entry into the joint settings tensor.
settings = jax.tree_map(
lambda o, v: jax.lax.dynamic_update_index_in_dim(o, v, i, axis=0),
settings,
new_settings,
)
# Get the settings for the batch entry from the joint settings tensor.
settings_slice = jax.tree_map(lambda t: jnp.expand_dims(t[i], axis=0), settings)
# Process the first n-1 tokens of the prompt.
lm_outputs = hk_forward(
jnp.expand_dims(prompt, 0),
memory=slice,
length=jnp.expand_dims(length, 0),
active=settings_slice.active,
)
# The forward pass doesn't correctly set the `step` counter inside the memory. Manually
# override it so `hk_forward` uses the correct context length in the next call.
slice = lm_outputs.model_state
slice = slice._replace(
layers=[l._replace(step=jnp.array([length])) for l in slice.layers]
)
# Sample the actual output token.
rng_ = jnp.expand_dims(rng_, 0)
new_output = sample_token(rng_, lm_outputs, settings_slice)
# Update the KV cache/memory.
slice = jax.tree_map(pad_to_max_len, slice)
memory = insert_slice(memory, slice, length, i)
rng = jnp.expand_dims(rng, 0)
rngs = jax.lax.dynamic_update_index_in_dim(rngs, rng, i, axis=0)
# Move the network outputs for this batch entry into the joint output tensor.
last_output = jax.tree_util.tree_map(
lambda last, new: jax.lax.dynamic_update_index_in_dim(last, new, i, axis=0),
last_output,
new_output,
)
return rngs, last_output, memory, settings
sample_step_ = hk.without_apply_rng(hk.transform(hk_sample_step))
prefill_memory_ = hk.without_apply_rng(hk.transform(hk_prefill_memory))
new_memory_ = hk.without_apply_rng(hk.transform(hk_new_memory))
forward_ = hk.without_apply_rng(hk.transform(hk_forward))
rng = jax.random.PRNGKey(42)
dummy_tokens = jnp.zeros((1, max_len), jnp.int32)
with runner.mesh:
shapes = jax.eval_shape(forward_.init, rng, dummy_tokens)
self.params_sharding = jax.tree_util.tree_map_with_path(
apply_rules(runner.model.partition_rules()),
shapes,
)
ds = P("data")
ms = runner.model.model.get_memory_sharding()
self.sample_step = pjit.pjit(
sample_step_.apply,
in_shardings=(self.params_sharding, None, ds, ms, None),
out_shardings=(None, ds, ms),
donate_argnums=3,
)
self.prefill_memory = pjit.pjit(
functools.partial(prefill_memory_.apply),
in_shardings=(
self.params_sharding,
None,
ms,
None,
ds,
None,
None,
None,
None,
None,
),
out_shardings=(None, ds, ms, None),
donate_argnums=(2,),
)
self.new_memory = pjit.pjit(
new_memory_.apply,
static_argnums=(1, 2),
out_shardings=ms,
)
def run(self):
"""Generator that accepts prompts."""
runner = self.runner
mesh = runner.mesh
max_len = runner.model.sequence_len
batch_size = runner.batch_size
params = self.params
rngs = jax.random.split(jax.random.PRNGKey(1), batch_size)
with mesh:
memory = self.new_memory(params, batch_size, max_len)
settings = SampleSettings(
temperature=np.zeros((batch_size,), dtype=np.float32),
nucleus_p=np.zeros((batch_size,), dtype=np.float32),
mask=np.ones((batch_size, self.vocab_size), dtype=np.int32),
active=np.zeros((batch_size), dtype=np.int32),
)
last_output = SampleOutput(
token_id=np.zeros((batch_size, 1), dtype=np.int32),
prob=np.zeros((batch_size, 1), dtype=jnp.bfloat16),
top_k_token_ids=np.zeros((batch_size, TOP_K), dtype=np.int32),
top_k_probs=np.zeros((batch_size, TOP_K), dtype=jnp.bfloat16),
)
prompt = np.array([300, 400, 500, 600, 600, 700, 800])
new_settings = SampleSettings(
temperature=np.float32(1),
nucleus_p=np.float32(1),
mask=np.ones((self.vocab_size,), dtype=np.int32),
active=np.zeros((), dtype=np.int32),
)
rng_seed = np.uint64(1)
for size in self.pad_sizes:
if size > runner.model.sequence_len:
break
logger.info("Precompile {}".format(size))
prompt_len = len(prompt)
prompt = pad_to_size(prompt, size)
rngs, last_output, memory, settings = self.prefill_memory(
params,
rngs,
memory,
settings,
last_output,
prompt,
prompt_len,
rng_seed,
new_settings,
0,
)
with runner.mesh:
logger.info("Compiling...")
rngs, last_output, memory = self.sample_step(
params, rngs, last_output, memory, settings
)
logger.info("Done compiling.")
all_tokens = []
free_slots = list(range(batch_size))
requests = [None] * batch_size
first_output = [None] * batch_size
jax.tree_map(lambda x: x.copy_to_host_async(), last_output)
prev_token = last_output
step = 0
total_num_tokens = 0
total_num_sequences = 0
with mesh:
while True:
while free_slots:
request: Optional[Request] = yield
tokens = self.tokenizer.encode(request.prompt)
temperature = request.temperature
nucleus_p = request.nucleus_p
rng_seed = request.rng_seed
i = free_slots.pop()
prompt = np.array(tokens, dtype=np.int32)
prompt_len = len(prompt)
prompt = pad_to_size(prompt, self.get_pad_bucket(prompt.shape[0]))
# All tokens are allowed.
mask = np.ones((self.vocab_size,), dtype=np.int32)
new_settings = SampleSettings(
temperature=np.float32(temperature),
nucleus_p=np.float32(nucleus_p),
mask=mask,
active=np.ones((), dtype=np.int32),
)
rng_seed = np.uint64(rng_seed)
rngs, last_output, memory, settings = self.prefill_memory(
params,
rngs,
memory,
settings,
last_output,
prompt,
prompt_len,
rng_seed,
new_settings,
i,
)
jax.tree_map(lambda x: x.copy_to_host_async(), last_output)
first_output[i] = last_output
requests[i] = request
total_num_sequences += 1
rngs, last_output, memory = self.sample_step(
params, rngs, last_output, memory, settings
)
total_num_tokens += batch_size - len(free_slots)
# prev_token should already be on the host.
prev_token = jax.tree_map(np.array, prev_token)
for i in range(batch_size):
if requests[i] is not None:
if first_output[i] is not None:
first_output_i = jax.tree_map(np.array, first_output[i])
all_tokens.append(int(first_output_i.token_id[i][0]))
first_output[i] = None
continue
all_tokens.append(int(prev_token.token_id[i][0]))
cont = len(all_tokens) < requests[i].max_len
if not cont:
output_str = self.tokenizer.decode(all_tokens)
requests[i] = None
free_slots.append(i)
all_tokens = []
settings = settings._replace(active=settings.active.at[i].set(0))
yield output_str
jax.tree_map(lambda x: x.copy_to_host_async(), last_output)
prev_token = last_output
step += 1
def make_mesh(
local_mesh_config: tuple[int, ...], between_hosts_config: tuple[int, ...]
) -> jax.sharding.Mesh:
assert len(local_mesh_config) == 2
assert len(between_hosts_config) == 2
rank_logger.info("Detected %s devices in mesh", jax.device_count())
device_mesh = mesh_utils.create_hybrid_device_mesh(
local_mesh_config,
between_hosts_config,
devices=jax.devices(),
process_is_granule=True,
)
rank_logger.debug(re.sub("\n+", "\n", f"Job device mesh is:\n{device_mesh}"))
return jax.sharding.Mesh(device_mesh, ("data", "model"))
def sample_from_model(server, prompt, max_len, temperature):
next(server)
inp = Request(
prompt=prompt,
temperature=temperature,
nucleus_p=1.0,
rng_seed=42,
max_len=max_len,
)
return server.send(inp)

BIN
tokenizer.model Normal file
View File

Binary file not shown.