broadcast grad acc fix to all models (#354)

## Summary
<!--- This is a required section; please describe the main purpose of
this proposed code change. --->

<!---
## Details
This is an optional section; is there anything specific that reviewers
should be aware of?
--->

follow up for #339

However, identify few issues
1. revert patching causes flce not taking effect (comment out revert
patching for now, and only test float32)
2. qwen2 vl flce is broken. we should fix later
3. we should provide a real "on-instance" patch that does not use any
monkey patch. now the on-instance patch still relies on monkey patch

## Testing Done
<!--- This is a required section; please describe how this change was
tested. --->

<!-- 
Replace BLANK with your device type. For example, A100-80G-PCIe

Complete the following tasks before sending your PR, and replace `[ ]`
with
`[x]` to indicate you have done them. 
-->

- Hardware Type: <BLANK>
- [ ] run `make test` to ensure correctness
- [ ] run `make checkstyle` to ensure code style
- [ ] run `make test-convergence` to ensure convergence

dev/modal/tests.py

@@ -17,6 +17,7 @@
 @app.function(gpu="A10G", mounts=[repo], timeout=60 * 10)
 def liger_tests():
     import subprocess
+
     subprocess.run(["pip", "install", "-e", "."], check=True, cwd="/root/liger-kernel")
     subprocess.run(["make", "test"], check=True, cwd="/root/liger-kernel")
     subprocess.run(["make", "test-convergence"], check=True, cwd="/root/liger-kernel")

src/liger_kernel/transformers/model/gemma.py

@@ -22,7 +22,7 @@
 @replace_return_docstrings(
     output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
 )
-def lce_forward(
+def lce_forward_deprecated(
     self,
     input_ids: torch.LongTensor = None,
     attention_mask: Optional[torch.Tensor] = None,
@@ -136,3 +136,126 @@ def lce_forward(
         hidden_states=outputs.hidden_states,
         attentions=outputs.attentions,
     )
+
+
+@add_start_docstrings_to_model_forward(GEMMA_INPUTS_DOCSTRING)
+@replace_return_docstrings(
+    output_type=CausalLMOutputWithPast, 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[Union[Cache, List[torch.FloatTensor]]] = 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,
+    return_dict: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    num_logits_to_keep: int = 0,
+    **loss_kwargs,
+) -> Union[Tuple, CausalLMOutputWithPast]:
+    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`, *optional*):
+            Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). 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 or large vocabulary size.
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, GemmaForCausalLM
+
+    >>> model = GemmaForCausalLM.from_pretrained("google/gemma-7b")
+    >>> tokenizer = AutoTokenizer.from_pretrained("google/gemma-7b")
+
+    >>> prompt = "What is your favorite condiment?"
+    >>> 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]
+    "What is your favorite condiment?"
+    ```"""
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    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,
+        return_dict=return_dict,
+        cache_position=cache_position,
+    )
+
+    hidden_states = outputs[0]
+
+    logits = None
+    loss = None
+    # if in training mode, don't materialize logits
+    if self.training and (labels is not None):
+        # We do the same thing as ForCausalLMLoss but using Liger FLCE
+
+        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)
+
+        reduction = "sum" if "num_items_in_batch" in loss_kwargs else "mean"
+        lce = LigerFusedLinearCrossEntropyLoss(reduction=reduction)
+
+        loss = lce(self.lm_head.weight, shift_hidden_states, shift_labels)
+        if reduction == "sum":
+            loss /= loss_kwargs["num_items_in_batch"]
+
+    else:  # if in inference mode materialize logits
+        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.vocab_size,
+                **loss_kwargs,
+            )
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+    )

src/liger_kernel/transformers/model/mistral.py

@@ -136,3 +136,6 @@ def lce_forward(
         hidden_states=outputs.hidden_states,
         attentions=outputs.attentions,
     )
+
+
+# Note: Grad Acc is not fixed in mistral at transformer 4.46.1

src/liger_kernel/transformers/model/mixtral.py

@@ -22,7 +22,7 @@
 @replace_return_docstrings(
     output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
 )
-def lce_forward(
+def lce_forward_deprecated(
     self,
     input_ids: torch.LongTensor = None,
     attention_mask: Optional[torch.Tensor] = None,
@@ -157,3 +157,153 @@ def lce_forward(
         attentions=outputs.attentions,
         router_logits=outputs.router_logits,
     )
+
+
+@add_start_docstrings_to_model_forward(MIXTRAL_INPUTS_DOCSTRING)
+@replace_return_docstrings(
+    output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
+)
+# Ignore copy
+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[List[torch.FloatTensor]] = 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: int = 0,
+    **loss_kwargs,
+) -> 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`, *optional*):
+            Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). 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 or large vocabulary size.
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, MixtralForCausalLM
+
+    >>> model = MixtralForCausalLM.from_pretrained("mistralai/Mixtral-8x7B-v0.1")
+    >>> tokenizer = AutoTokenizer.from_pretrained("mistralai/Mixtral-8x7B-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,
+        return_dict=return_dict,
+        cache_position=cache_position,
+    )
+
+    hidden_states = outputs[0]
+
+    logits = None
+    loss = None
+    # if in training mode, don't materialize logits
+    if self.training and (labels is not None):
+        # We do the same thing as ForCausalLMLoss but using Liger FLCE
+
+        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)
+
+        reduction = "sum" if "num_items_in_batch" in loss_kwargs else "mean"
+        lce = LigerFusedLinearCrossEntropyLoss(reduction=reduction)
+
+        loss = lce(self.lm_head.weight, shift_hidden_states, shift_labels)
+        if reduction == "sum":
+            loss /= loss_kwargs["num_items_in_batch"]
+
+    else:  # if in inference mode materialize logits
+        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.vocab_size,
+                **loss_kwargs,
+            )
+
+    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,
+    )