Add support for jamba model with Liger Kernel

pyproject.toml

@@ -1,5 +1,5 @@
 [build-system]
-requires = ["setuptools>=42", "wheel"]
+requires = ["setuptools>=61.0", "wheel"]
 build-backend = "setuptools.build_meta"
 
 [project]
@@ -27,6 +27,11 @@ dev = [
     "seaborn",
 ]
 
+test = [
+    "causal-conv1d>=1.4.0",
+    "mamba-ssm>=2.2.2",
+]
+
 [tool.setuptools.packages.find]
 where = ["src"]
 include = ["liger_kernel", "liger_kernel.*"]

src/liger_kernel/transformers/__init__.py

@@ -10,6 +10,7 @@
 from liger_kernel.transformers.monkey_patch import (  # noqa: F401
     apply_liger_kernel_to_gemma,
     apply_liger_kernel_to_gemma2,
+    apply_liger_kernel_to_jamba,
     apply_liger_kernel_to_llama,
     apply_liger_kernel_to_mistral,
     apply_liger_kernel_to_mixtral,

src/liger_kernel/transformers/model/jamba.py

@@ -0,0 +1,169 @@
+from typing import Optional, Tuple, Union
+
+import torch
+from torch.nn import CrossEntropyLoss
+from transformers.modeling_outputs import MoeCausalLMOutputWithPast
+from transformers.models.jamba.modeling_jamba import (
+    _CONFIG_FOR_DOC,
+    JAMBA_INPUTS_DOCSTRING,
+    HybridMambaAttentionDynamicCache,
+    load_balancing_loss_func,
+)
+from transformers.utils import (
+    add_start_docstrings_to_model_forward,
+    replace_return_docstrings,
+)
+
+from liger_kernel.transformers.fused_linear_cross_entropy import (
+    LigerFusedLinearCrossEntropyLoss,
+)
+
+
+@add_start_docstrings_to_model_forward(JAMBA_INPUTS_DOCSTRING)
+@replace_return_docstrings(
+    output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
+)
+def lce_forward(
+    self,
+    input_ids: torch.LongTensor = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[HybridMambaAttentionDynamicCache] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    output_router_logits: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    num_logits_to_keep: Optional[Union[int, None]] = None,
+) -> Union[Tuple, MoeCausalLMOutputWithPast]:
+    r"""
+    Args:
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        num_logits_to_keep (`int` or `None`, *optional*):
+            Calculate logits for the last `num_logits_to_keep` tokens. If `None`, calculate logits for all
+            `input_ids`. Only last token logits are needed for generation, and calculating them only for that token
+            can save memory, which becomes pretty significant for long sequences.
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, JambaForCausalLM
+
+    >>> model = JambaForCausalLM.from_pretrained("ai21labs/Jamba-v0.1")
+    >>> tokenizer = AutoTokenizer.from_pretrained("ai21labs/Jamba-v0.1")
+
+    >>> prompt = "Hey, are you conscious? Can you talk to me?"
+    >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+    >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+    ```"""
+
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_router_logits = (
+        output_router_logits
+        if output_router_logits is not None
+        else self.config.output_router_logits
+    )
+
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+    return_dict = (
+        return_dict if return_dict is not None else self.config.use_return_dict
+    )
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs = self.model(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        output_router_logits=output_router_logits,
+        cache_position=cache_position,
+        return_dict=return_dict,
+    )
+
+    hidden_states = outputs[0]
+
+    loss = None
+    logits = None
+
+    if self.training:
+        shift_hidden_states = hidden_states[..., :-1, :].contiguous()
+        shift_labels = labels[..., 1:].contiguous()
+
+        # flatten tokens
+        shift_hidden_states = shift_hidden_states.view(-1, self.config.hidden_size)
+        shift_labels = shift_labels.view(-1)
+
+        lce = LigerFusedLinearCrossEntropyLoss()
+        loss = lce(self.lm_head.weight, shift_hidden_states, shift_labels)
+    else:
+        if num_logits_to_keep is None:
+            logits = self.lm_head(hidden_states)
+        else:
+            logits = self.lm_head(hidden_states[..., -num_logits_to_keep:, :])
+        logits = logits.float()
+
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+    aux_loss = None
+    if output_router_logits:
+        aux_loss = load_balancing_loss_func(
+            outputs.router_logits if return_dict else outputs[-1],
+            self.num_experts,
+            self.num_experts_per_tok,
+            attention_mask,
+        )
+        if labels is not None:
+            loss += self.router_aux_loss_coef * aux_loss.to(
+                loss.device
+            )  # make sure to reside in the same device
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        if output_router_logits:
+            output = (aux_loss,) + output
+        return (loss,) + output if loss is not None else output
+
+    return MoeCausalLMOutputWithPast(
+        loss=loss,
+        aux_loss=aux_loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        router_logits=outputs.router_logits,
+    )

src/liger_kernel/transformers/monkey_patch.py

@@ -6,6 +6,7 @@
 from liger_kernel.transformers.geglu import LigerGEGLUMLP
 from liger_kernel.transformers.layer_norm import LigerLayerNorm
 from liger_kernel.transformers.model.gemma import lce_forward as gemma_lce_forward
+from liger_kernel.transformers.model.jamba import lce_forward as jamba_lce_forward
 from liger_kernel.transformers.model.llama import lce_forward as llama_lce_forward
 from liger_kernel.transformers.model.mistral import lce_forward as mistral_lce_forward
 from liger_kernel.transformers.model.mixtral import lce_forward as mixtral_lce_forward
@@ -332,6 +333,43 @@ def apply_liger_kernel_to_phi3(
         modeling_phi3.Phi3ForCausalLM.forward = phi3_lce_forward
 
 
+def apply_liger_kernel_to_jamba(
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = True,
+    rms_norm: bool = True,
+    swiglu: bool = True,
+) -> None:
+    """
+    Apply Liger kernels to replace original implementation in HuggingFace Jamba models
+    to make GPU go burrr.
+
+    # Note: Jamba model does not use rotary position embedding(RoPE).
+
+    Args:
+        cross_entropy (bool): Whether to apply Liger's cross entropy loss. Default is False.
+        fused_linear_cross_entropy (bool):
+            Whether to apply Liger's fused lienar cross entropy loss. Default is True.
+            `cross_entropy` and `fused_linear_cross_entropy` cannot both be True.
+            If `fused_linear_cross_entropy` is True, the logits will not be materialized but more memory efficient.
+        rms_norm (bool): Whether to apply Liger's RMSNorm. Default is True.
+        geglu (bool): Whether to apply Liger's GeGLU MLP. Default is True.
+    """
+    assert not (
+        cross_entropy and fused_linear_cross_entropy
+    ), "cross_entropy and fused_linear_cross_entropy cannot both be True."
+
+    from transformers.models.jamba import modeling_jamba
+
+    if rms_norm:
+        modeling_jamba.JambaRMSNorm = LigerRMSNorm
+    if cross_entropy:
+        modeling_jamba.CrossEntropyLoss = LigerCrossEntropyLoss
+    if swiglu:
+        modeling_jamba.JambaMLP = LigerSwiGLUMLP
+    if fused_linear_cross_entropy:
+        modeling_jamba.JambaForCausalLM.forward = jamba_lce_forward
+
+
 # Model type corresponds to the keys defined in transformers/models/auto/modeling_auto.py
 MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "gemma": apply_liger_kernel_to_gemma,
@@ -342,6 +380,7 @@ def apply_liger_kernel_to_phi3(
     "qwen2": apply_liger_kernel_to_qwen2,
     "qwen2_vl": apply_liger_kernel_to_qwen2_vl,
     "phi3": apply_liger_kernel_to_phi3,
+    "jamba": apply_liger_kernel_to_jamba,
 }
 
 

test/convergence/test_mini_models.py

@@ -15,6 +15,7 @@
 from torch.utils.data import DataLoader
 from transformers.models.gemma import GemmaConfig, GemmaForCausalLM
 from transformers.models.gemma2 import Gemma2Config, Gemma2ForCausalLM
+from transformers.models.jamba import JambaConfig, JambaForCausalLM
 from transformers.models.llama import LlamaConfig, LlamaForCausalLM
 from transformers.models.mistral import MistralConfig, MistralForCausalLM
 from transformers.models.mixtral import MixtralConfig, MixtralForCausalLM
@@ -24,6 +25,7 @@
 from liger_kernel.transformers import (
     apply_liger_kernel_to_gemma,
     apply_liger_kernel_to_gemma2,
+    apply_liger_kernel_to_jamba,
     apply_liger_kernel_to_llama,
     apply_liger_kernel_to_mistral,
     apply_liger_kernel_to_mixtral,
@@ -279,6 +281,31 @@
             attn_implementation="eager",
         ),
     ),
+    "mini_jamba": MiniModelConfig(
+        liger_kernel_patch_func=functools.partial(
+            apply_liger_kernel_to_jamba, fused_linear_cross_entropy=False
+        ),
+        model_class=JambaForCausalLM,
+        mini_model_config=JambaConfig(
+            attention_dropout=0.0,
+            num_experts_per_tok=1,
+            num_experts=2,
+            bos_token_id=1,
+            eos_token_id=2,  # 32000
+            hidden_act="silu",
+            hidden_size=1024,  # 3072
+            initializer_range=0.02,
+            intermediate_size=2048,  # 8192
+            max_position_embeddings=32768,
+            num_attention_heads=8,  # 32
+            num_hidden_layers=4,  # 32
+            rms_norm_eps=1e-5,
+            sliding_window=None,
+            tie_word_embeddings=False,
+            use_cache=True,
+            vocab_size=32064,
+        ),
+    ),
 }
 
 
@@ -316,6 +343,8 @@ def run_mini_model(
             kwargs["geglu"] = True
         else:
             kwargs["swiglu"] = True
+        if model_name == "mini_jamba":
+            del kwargs["rope"]
         MINI_MODEL_SETUPS[model_name].liger_kernel_patch_func(**kwargs)
 
     model = create_model(model_name).to(dtype).to("cuda")
@@ -328,7 +357,6 @@ def run_mini_model(
     optimizer = torch.optim.AdamW(model.parameters(), lr=lr)
 
     loss_list = []
-
     for i in range(num_steps):
         batch = next(loader_iter).to(model.device)
         optimizer.zero_grad()
@@ -461,6 +489,8 @@ def run_mini_model(
                 not supports_bfloat16(), reason="bfloat16 not supported on this GPU"
             ),
         ),
+        # To run this test, you need to first run `pip install . '[test]'`
+        # ("mini_jamba", 32, 1e-4, torch.float32, 1e-8, 1e-5, 5e-3, 1e-5, 5e-3, 1e-5),
     ],
 )
 def test_mini_model(

test/convergence/test_mini_models_no_logits.py

@@ -11,6 +11,7 @@
 import torch
 from datasets import load_from_disk
 from torch.utils.data import DataLoader
+from transformers import JambaConfig, JambaForCausalLM
 from transformers.models.gemma import GemmaConfig, GemmaForCausalLM
 from transformers.models.gemma2 import Gemma2Config, Gemma2ForCausalLM
 from transformers.models.llama import LlamaConfig, LlamaForCausalLM
@@ -22,6 +23,7 @@
 from liger_kernel.transformers import (
     apply_liger_kernel_to_gemma,
     apply_liger_kernel_to_gemma2,
+    apply_liger_kernel_to_jamba,
     apply_liger_kernel_to_llama,
     apply_liger_kernel_to_mistral,
     apply_liger_kernel_to_mixtral,
@@ -259,6 +261,29 @@
             attention_dropout=0.0,
         ),
     ),
+    "mini_jamba": MiniModelConfig(
+        liger_kernel_patch_func=apply_liger_kernel_to_jamba,
+        model_class=JambaForCausalLM,
+        mini_model_config=JambaConfig(
+            attention_dropout=0.0,
+            num_experts_per_tok=1,
+            num_experts=2,
+            bos_token_id=1,
+            eos_token_id=2,  # 32000
+            hidden_act="silu",
+            hidden_size=1024,  # 3072
+            initializer_range=0.02,
+            intermediate_size=2048,  # 8192
+            max_position_embeddings=32768,
+            num_attention_heads=8,  # 32
+            num_hidden_layers=4,  # 32
+            rms_norm_eps=1e-5,
+            sliding_window=None,
+            tie_word_embeddings=False,
+            use_cache=True,
+            vocab_size=32064,
+        ),
+    ),
 }
 
 if QWEN2_VL_AVAILABLE:
@@ -346,7 +371,8 @@ def run_mini_model(
             kwargs["geglu"] = True
         else:
             kwargs["swiglu"] = True
-
+        if model_name == "mini_jamba":
+            del kwargs["rope"]
         model_support_flce = "gemma2" not in model_name
         if model_support_flce:
             kwargs["fused_linear_cross_entropy"] = True
@@ -547,6 +573,8 @@ def run_mini_model(
                 not supports_bfloat16(), reason="bfloat16 not supported on this GPU"
             ),
         ),
+        # To run this test, you need to first run `pip install . '[test]'`
+        # ("mini_jamba", 32, 1e-4, torch.float32, 5e-4, 1e-4, 5e-3, 1e-5, 1e-2, 1e-5),
     ],
 )
 def test_mini_model(