Add codestral mamba2 (#32080)

* add new model like * draft cuda forward - mismatched keys (sharding on conv1) * match keys successfully * fix split * get generation/forward running (wrong gens, norm?) * :update * some refactoring * fixes * works up until copy to cache * fix * update * NON WORKING VERSION * version that work? * nit * fix config * fix conversion script * working cuda forward * nit * update * simplifcation * make mamba slow simple work * no einops * todo * fix style * no einops * update fix no einsum * nit * remove einops * bug: scan_output differs strongly * add rms norm option * fix fast + slow generation with and w/o cache ✔️ * draft integration tests * remove a big chunk of the einsum * fix slow, fast generations, without any einsum * fix copies * fix structure * fix up modeling and tests * fix tests * clamping is indeed worse * recover mamba2 cache test * fix copies * no cache position (yet) * fix tf tests * fix matmul for generate * fixup * skip cache tests for now * [run-slow]mamba2 * tune out hidden states for padding * test batched generation * propagate attention mask changes * fix past length * fix integration test * style * address comments * update readme * add mamba2 version check * fix tests * [run-slow]mamba2 * skip edge tests * [run-slow]mamba2 * last fixup * [run-slow]mamba2 * update README --------- Co-authored-by: Arthur Zucker <arthur.zucker@gmail.com>
huggingface · Aug 6, 2024 · 80b90e7 · 80b90e7
1 parent 3d8bd11
commit 80b90e7
Show file tree

Hide file tree

Showing 16 changed files with 1,947 additions and 0 deletions.
diff --git a/docs/source/en/_toctree.yml b/docs/source/en/_toctree.yml
@@ -438,6 +438,8 @@
         title: MADLAD-400
       - local: model_doc/mamba
         title: Mamba
+      - local: model_doc/mamba2
+        title: mamba2
       - local: model_doc/marian
         title: MarianMT
       - local: model_doc/markuplm

diff --git a/docs/source/en/index.md b/docs/source/en/index.md
@@ -194,6 +194,7 @@ Flax), PyTorch, and/or TensorFlow.
 |                       [M2M100](model_doc/m2m_100)                        |       ✅        |         ❌         |      ❌      |
 |                    [MADLAD-400](model_doc/madlad-400)                    |       ✅        |         ✅         |      ✅      |
 |                         [Mamba](model_doc/mamba)                         |       ✅        |         ❌         |      ❌      |
+|                        [mamba2](model_doc/mamba2)                        |       ✅        |         ❌         |      ❌      |
 |                        [Marian](model_doc/marian)                        |       ✅        |         ✅         |      ✅      |
 |                      [MarkupLM](model_doc/markuplm)                      |       ✅        |         ❌         |      ❌      |
 |                   [Mask2Former](model_doc/mask2former)                   |       ✅        |         ❌         |      ❌      |

diff --git a/docs/source/en/model_doc/mamba2.md b/docs/source/en/model_doc/mamba2.md
@@ -0,0 +1,106 @@
+<!--Copyright 2024 The HuggingFace 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.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Mamba 2
+
+## Overview
+
+The Mamba2 model was proposed in [Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality](https://arxiv.org/abs/2405.21060) by Tri Dao and Albert Gu. It is a State Space Model similar to Mamba 1, with better performances in a simplified architecture. 
+
+
+The abstract from the paper is the following:
+
+*While Transformers have been the main architecture behind deep learning's success in language modeling, state-space models (SSMs) such as Mamba have recently been shown to match or outperform Transformers at small to medium scale. We show that these families of models are actually quite closely related, and develop a rich framework of theoretical connections between SSMs and variants of attention, connected through various decompositions of a well-studied class of structured semiseparable matrices. Our state space duality (SSD) framework allows us to design a new architecture (Mamba-2) whose core layer is an a refinement of Mamba's selective SSM that is 2-8X faster, while continuing to be competitive with Transformers on language modeling.*
+
+Tips:
+
+This version should support all implementations of Mamba 2, and in particular [Mamba-2 codestral](https://huggingface.co/mistralai/Mamba-Codestral-7B-v0.1) from Mistral AI. In particular, mamba 2 codestral was released with a number of `groups` equal to 8, which can be thought intuitively as similar to the number of kv heads in an attention-based model. 
+This model has two different forward passes, `torch_forward` or `cuda_kernels_forward`. The latter uses the original cuda kernels if they are found in your environment, and is slower on the prefill i.e. requires a "warmup run" due to high cpu overhead, see [here](https://github.com/state-spaces/mamba/issues/389#issuecomment-2171755306) and [also here](https://github.com/state-spaces/mamba/issues/355#issuecomment-2147597457). Without compilation, the `torch_forward` implementation is faster by a factor 3 to 4. Further, there are no positional embeddings in this model, but there is an `attention_mask` and a specific logic to mask out hidden states in two places in the case of batched generation, see [here](https://github.com/state-spaces/mamba/issues/66#issuecomment-1863563829) as well. Due to this, in addition to the reimplementation of mamba2 kernels, batched generation and cached generation are expected to have slight discrepancies. Further, the results given by the cuda kernels or the torch forward are expected to be slightly different. The SSM algorithm heavily relies on tensor contractions, which have matmul equivalents but the order of operations is slightly different, making the difference greater at smaller precisions. 
+Another note, shutdown of hidden states corresponding to padding tokens is done in 2 places and mostly has been tested with left-padding. Right-padding will propagate noise down the line and is not guaranteed to yield satisfactory results. `tokenizer.padding_side = "left"` ensures you are using the correct padding side.
+
+This model was contributed by [Molbap](https://huggingface.co/Molbap), with tremendous help from [Anton Vlasjuk](https://github.com/vasqu).
+The original code can be found [here](https://github.com/state-spaces/mamba).
+
+
+# Usage
+
+### A simple generation example: 
+```python 
+from transformers import MambaConfig, MambaForCausalLM, AutoTokenizer
+import torch
+model_id = 'mistralai/Mamba-Codestral-7B-v0.1'
+tokenizer = AutoTokenizer.from_pretrained(model_id, revision='refs/pr/9', from_slow=True, legacy=False)
+model = MambaForCausalLM.from_pretrained(model_id, revision='refs/pr/9')
+input_ids = tokenizer("Hey how are you doing?", return_tensors= "pt")["input_ids"]
+
+out = model.generate(input_ids, max_new_tokens=10)
+print(tokenizer.batch_decode(out))
+```
+
+Here's a draft script for finetuning: 
+```python 
+from trl import SFTTrainer
+from peft import LoraConfig
+from transformers import AutoTokenizer, Mamba2ForCausalLM, TrainingArguments
+model_id = 'mistralai/Mamba-Codestral-7B-v0.1'
+tokenizer = AutoTokenizer.from_pretrained(model_id, revision='refs/pr/9', from_slow=True, legacy=False)
+tokenizer.pad_token = tokenizer.eos_token
+tokenizer.padding_side = "left" #enforce padding side left
+
+model = Mamba2ForCausalLM.from_pretrained(model_id, revision='refs/pr/9')
+dataset = load_dataset("Abirate/english_quotes", split="train")
+# Without CUDA kernels, batch size of 2 occupies one 80GB device
+# but precision can be reduced.
+# Experiments and trials welcome!
+training_args = TrainingArguments(
+    output_dir="./results",
+    num_train_epochs=3,
+    per_device_train_batch_size=2,
+    logging_dir='./logs',
+    logging_steps=10,
+    learning_rate=2e-3
+)
+lora_config =  LoraConfig(
+        r=8,
+        target_modules=["embeddings", "in_proj", "out_proj"],
+        task_type="CAUSAL_LM",
+        bias="none"
+)
+trainer = SFTTrainer(
+    model=model,
+    tokenizer=tokenizer,
+    args=training_args,
+    peft_config=lora_config,
+    train_dataset=dataset,
+    dataset_text_field="quote",
+)
+trainer.train()
+```
+
+
+## Mamba2Config
+
+[[autodoc]] Mamba2Config
+
+## Mamba2Model
+
+[[autodoc]] Mamba2Model
+    - forward
+
+## Mamba2LMHeadModel
+
+[[autodoc]] Mamba2ForCausalLM
+    - forward
diff --git a/src/transformers/__init__.py b/src/transformers/__init__.py
@@ -544,6 +544,7 @@
     ],
     "models.m2m_100": ["M2M100Config"],
     "models.mamba": ["MambaConfig"],
+    "models.mamba2": ["Mamba2Config"],
     "models.marian": ["MarianConfig"],
     "models.markuplm": [
         "MarkupLMConfig",
@@ -2550,6 +2551,13 @@
             "MambaPreTrainedModel",
         ]
     )
+    _import_structure["models.mamba2"].extend(
+        [
+            "Mamba2ForCausalLM",
+            "Mamba2Model",
+            "Mamba2PreTrainedModel",
+        ]
+    )
     _import_structure["models.marian"].extend(["MarianForCausalLM", "MarianModel", "MarianMTModel"])
     _import_structure["models.markuplm"].extend(
         [
@@ -5240,6 +5248,7 @@
     )
     from .models.m2m_100 import M2M100Config
     from .models.mamba import MambaConfig
+    from .models.mamba2 import Mamba2Config
     from .models.marian import MarianConfig
     from .models.markuplm import (
         MarkupLMConfig,
@@ -7046,6 +7055,11 @@
             MambaModel,
             MambaPreTrainedModel,
         )
+        from .models.mamba2 import (
+            Mamba2ForCausalLM,
+            Mamba2Model,
+            Mamba2PreTrainedModel,
+        )
         from .models.marian import MarianForCausalLM, MarianModel, MarianMTModel
         from .models.markuplm import (
             MarkupLMForQuestionAnswering,

diff --git a/src/transformers/models/__init__.py b/src/transformers/models/__init__.py
@@ -135,6 +135,7 @@
     lxmert,
     m2m_100,
     mamba,
+    mamba2,
     marian,
     markuplm,
     mask2former,

diff --git a/src/transformers/models/auto/configuration_auto.py b/src/transformers/models/auto/configuration_auto.py
@@ -152,6 +152,7 @@
         ("lxmert", "LxmertConfig"),
         ("m2m_100", "M2M100Config"),
         ("mamba", "MambaConfig"),
+        ("mamba2", "Mamba2Config"),
         ("marian", "MarianConfig"),
         ("markuplm", "MarkupLMConfig"),
         ("mask2former", "Mask2FormerConfig"),
@@ -440,6 +441,7 @@
         ("m2m_100", "M2M100"),
         ("madlad-400", "MADLAD-400"),
         ("mamba", "Mamba"),
+        ("mamba2", "mamba2"),
         ("marian", "Marian"),
         ("markuplm", "MarkupLM"),
         ("mask2former", "Mask2Former"),

diff --git a/src/transformers/models/auto/modeling_auto.py b/src/transformers/models/auto/modeling_auto.py
@@ -144,6 +144,7 @@
         ("lxmert", "LxmertModel"),
         ("m2m_100", "M2M100Model"),
         ("mamba", "MambaModel"),
+        ("mamba2", "Mamba2Model"),
         ("marian", "MarianModel"),
         ("markuplm", "MarkupLMModel"),
         ("mask2former", "Mask2FormerModel"),
@@ -310,6 +311,7 @@
         ("luke", "LukeForMaskedLM"),
         ("lxmert", "LxmertForPreTraining"),
         ("mamba", "MambaForCausalLM"),
+        ("mamba2", "Mamba2ForCausalLM"),
         ("mega", "MegaForMaskedLM"),
         ("megatron-bert", "MegatronBertForPreTraining"),
         ("mobilebert", "MobileBertForPreTraining"),
@@ -394,6 +396,7 @@
         ("luke", "LukeForMaskedLM"),
         ("m2m_100", "M2M100ForConditionalGeneration"),
         ("mamba", "MambaForCausalLM"),
+        ("mamba2", "Mamba2ForCausalLM"),
         ("marian", "MarianMTModel"),
         ("mega", "MegaForMaskedLM"),
         ("megatron-bert", "MegatronBertForCausalLM"),
@@ -472,6 +475,7 @@
         ("jetmoe", "JetMoeForCausalLM"),
         ("llama", "LlamaForCausalLM"),
         ("mamba", "MambaForCausalLM"),
+        ("mamba2", "Mamba2ForCausalLM"),
         ("marian", "MarianForCausalLM"),
         ("mbart", "MBartForCausalLM"),
         ("mega", "MegaForCausalLM"),

diff --git a/src/transformers/models/auto/tokenization_auto.py b/src/transformers/models/auto/tokenization_auto.py
@@ -270,6 +270,7 @@
             ("lxmert", ("LxmertTokenizer", "LxmertTokenizerFast" if is_tokenizers_available() else None)),
             ("m2m_100", ("M2M100Tokenizer" if is_sentencepiece_available() else None, None)),
             ("mamba", (None, "GPTNeoXTokenizerFast" if is_tokenizers_available() else None)),
+            ("mamba2", (None, "GPTNeoXTokenizerFast" if is_tokenizers_available() else None)),
             ("marian", ("MarianTokenizer" if is_sentencepiece_available() else None, None)),
             (
                 "mbart",

diff --git a/src/transformers/models/mamba2/__init__.py b/src/transformers/models/mamba2/__init__.py
@@ -0,0 +1,58 @@
+# Copyright 2024 The HuggingFace 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_mamba2": ["Mamba2Config", "Mamba2OnnxConfig"],
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_mamba2"] = [
+        "Mamba2ForCausalLM",
+        "Mamba2Model",
+        "Mamba2PreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_mamba2 import Mamba2Config, Mamba2OnnxConfig
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_mamba2 import (
+            Mamba2ForCausalLM,
+            Mamba2Model,
+            Mamba2PreTrainedModel,
+        )
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)