Add the Bamba Model

[fabianlim]

What does this PR do?

This PR merges the BambaModel, which is a hybrid mamba2 architecture with SwiGLU. The checkpoints are jointly trained by IBM, Princeton, and UIUC.

The implementation is based off ai21labs/Jamba-v0.1 and the mamba2 implementation ported over to HF for the codestral model.

cc: @ani300, @raghukiran1224

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[Rocketknight1]

Hi @fabianlim, do you have a paper reference for this model or any details on the trained checkpoints?

[fabianlim]

@Rocketknight1 thanks for reaching out. Yes my colleagues are preparing a paper and a GitHub repo with the (training) code. And checkpoints will be 1.8T, 2T, 2.2T, and an sft model. We will update the PR accordingly.

cc: @raghukiran1224

[raghukiran1224]

The data used is all open, we plan to share any and all details on what the community would want! Open source is the name of the game 😄

[Rocketknight1]

Cool! @molbap will be the point of contact at Hugging Face for this PR, so feel free to ping me or him if you have any questions as you're working on it.

[molbap]

All tests + integration pass and modular looks better, thanks for working on it and congrats on the model! pinging @ArthurZucker so core review starts ;)

[molbap]

This should be filled in as well before merging

[fabianlim]

updated in 44788dc

[molbap]

Suggested change

<!-- update this -->

[fabianlim]

updated in 44788dc

[ArthurZucker]

Great work, a few nits to adresse and we can merge!

[ArthurZucker]

Suggested change

	# Copyright 2024 IBM and the HuggingFace Inc. team. All rights reserved.
	#
	# 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 typing import TYPE_CHECKING
	from ...utils import (
	OptionalDependencyNotAvailable,
	_LazyModule,
	is_torch_available,
	)
	_import_structure = {
	"configuration_bamba": ["BambaConfig"],
	}
	try:
	if not is_torch_available():
	raise OptionalDependencyNotAvailable()
	except OptionalDependencyNotAvailable:
	pass
	else:
	_import_structure["modeling_bamba"] = [
	"BambaForCausalLM",
	"BambaModel",
	"BambaPreTrainedModel",
	]
	if TYPE_CHECKING:
	from .configuration_bamba import BambaConfig
	try:
	if not is_torch_available():
	raise OptionalDependencyNotAvailable()
	except OptionalDependencyNotAvailable:
	pass
	else:
	from .modeling_bamba import (
	BambaForCausalLM,
	BambaModel,
	BambaPreTrainedModel,
	)
	else:
	import sys
	sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
	# Copyright 2024 IBM and the HuggingFace Inc. team. All rights reserved.
	#
	# 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 typing import TYPE_CHECKING
	from ...utils import _LazyModule
	from ...utils.import_utils import define_import_structure
	if TYPE_CHECKING:
	from .configuration_bamba import *
	from .modeling_bamba import *
	from .processing_bamba import *
	else:
	import sys
	_file = globals()["__file__"]
	sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

[ArthurZucker]

and you need to define __all__ like:

__all__ = [
    "GemmaModel",
    "GemmaForCausalLM",
    "GemmaForSequenceClassification",
    "GemmaForTokenClassification",
    "GemmaPreTrainedModel",
]

at the end of the modular file! see modular gemma2

[ani300]

addressed in latest commit

[ArthurZucker]

let's remove asserts and rather raise errors please

[ani300]

done in latest commit

[ArthurZucker]

we like to have a more explicit dict like this one

transformers/src/transformers/models/mllama/convert_mllama_weights_to_hf.py

Line 41 in deac971

ORIGINAL_TO_CONVERTED_KEY_MAPPING = {

, but it's not blocking merge!

[ani300]

leaving for future/followup PR

[ArthurZucker]

really not sure this is worth changing as it's just ads an assert.

[ArthurZucker]

Otherwise we should use the notation like

transformers/src/transformers/models/glm/modular_glm.py

Line 71 in deac971

def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):

[ani300]

changed to adapt from GLM

[ArthurZucker]

as a reviewer, but also any dev that is gonna read this code, we need to know what the differences are with Mamba2Mixer.

Could you add comments here explaining why we have to redefine everything?

[ani300]

added a list to the comments

[ArthurZucker]

This is a lot more similar to :

class JambaAttentionDecoderLayer(nn.Module):
    def __init__(self, config: JambaConfig, layer_idx: int):
        super().__init__()
        num_experts = config.layers_num_experts[layer_idx]
        self.self_attn = JAMBA_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)

        ffn_layer_class = JambaSparseMoeBlock if num_experts > 1 else JambaMLP
        self.feed_forward = ffn_layer_class(config)
        self.input_layernorm = JambaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.pre_ff_layernorm = JambaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

which should be a better base 😉

[ani300]

done in latest commit

[ArthurZucker]

forward and init should be the same as LlamaForcausalLM which you should be able to inherit from and just change the prepare input for generation as it is the only difference no?

[ani300]

done in latest commit

[ArthurZucker]

Kudos! 🚀