cciwon-code-review-cli 2.0.1 → 2.0.3

package/unsloth_compiled_cache/unsloth_compiled_module_qwen3_moe.py

@@ -0,0 +1,726 @@
+"""
+2025.12.6
+2025.12.7
+4.57.1
+0.24.0
+__UNSLOTH_VERSIONING__
+"""
+
+# Unsloth auto generated code
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU Lesser General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+
+import os
+import torch
+import importlib.util
+import math
+if importlib.util.find_spec("unsloth_studio") is None:
+    UNSLOTH_STUDIO_ENABLED = False
+else:
+    UNSLOTH_STUDIO_ENABLED = os.environ.get("UNSLOTH_STUDIO_DISABLED", "0") == "0"
+pass
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+import math
+
+UNSLOTH_ENABLE_LOGGING = os.environ.get("UNSLOTH_ENABLE_LOGGING", "0") == "1"
+UNSLOTH_ENABLE_CCE = os.environ.get("UNSLOTH_ENABLE_CCE", "1") == "1"
+UNSLOTH_COMPILE_DISABLE = os.environ.get("UNSLOTH_COMPILE_DISABLE", "0") in ("1", "partial",)
+
+import logging
+logger_compiler = logging.getLogger(__name__)
+if UNSLOTH_ENABLE_LOGGING:
+    logger_compiler.setLevel(logging.DEBUG)
+
+global INFERENCE_RUNS
+INFERENCE_RUNS = 0
+
+try:
+    import torch._dynamo.eval_frame as torch_dynamo_eval_frame
+    torch_dynamo_eval_frame._stance.stance
+    torch_compiler_set_stance = torch.compiler.set_stance
+except:
+    torch_dynamo_eval_frame = None
+    torch_compiler_set_stance = None
+pass
+
+from unsloth_zoo import DEVICE_TYPE_TORCH, DEVICE_COUNT
+
+
+from unsloth_zoo.loss_utils import (
+    fused_linear_cross_entropy,
+    unsloth_fused_ce_loss,
+)
+
+if UNSLOTH_STUDIO_ENABLED:
+    from unsloth_zoo.loss_utils import fast_linear_cross_entropy
+
+scaled_dot_product_attention = torch.nn.functional.scaled_dot_product_attention
+@torch.compiler.disable(recursive = False)
+def disable_compile_scaled_dot_product_attention(*args, **kwargs):
+    return scaled_dot_product_attention(*args, **kwargs)
+pass
+
+
+from transformers.modeling_flash_attention_utils import is_flash_attn_available
+
+if is_flash_attn_available():
+    try:
+        from transformers.modeling_flash_attention_utils import flash_attn_supports_top_left_mask
+    except:
+        flash_attn_supports_top_left_mask = None
+    try:
+        from transformers.modeling_flash_attention_utils import _flash_attention_forward
+    except:
+        _flash_attention_forward = None
+    try:
+        from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+    except:
+        FlashAttentionKwargs = None
+    try:
+        from transformers.modeling_flash_attention_utils import flash_attn_varlen_func
+    except:
+        flash_attn_varlen_func = None
+else:
+    flash_attn_supports_top_left_mask = None
+    _flash_attention_forward = None
+    FlashAttentionKwargs = None
+    flash_attn_varlen_func = None
+pass
+
+
+torch_compile_options = {'epilogue_fusion': True, 'max_autotune': False, 'shape_padding': True, 'trace.enabled': False, 'triton.cudagraphs': False, 'debug': False, 'dce': True, 'memory_planning': True, 'coordinate_descent_tuning': False, 'trace.graph_diagram': False, 'compile_threads': 32, 'group_fusion': True, 'disable_progress': True, 'verbose_progress': False, 'triton.multi_kernel': 0, 'triton.use_block_ptr': False, 'triton.enable_persistent_tma_matmul': True, 'triton.autotune_at_compile_time': False, 'triton.cooperative_reductions': False, 'cuda.compile_opt_level': '-O2', 'cuda.enable_cuda_lto': True, 'combo_kernels': False, 'benchmark_combo_kernel': True, 'combo_kernel_foreach_dynamic_shapes': True}
+
+from torch.nn import CrossEntropyLoss
+
+@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
+def normal_cross_entropy_loss(self, hidden_states, labels):
+    logits = self.lm_head(hidden_states)
+    logits = logits.float()
+    # 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)
+    return loss, logits
+pass
+
+# We need an empty logits flag to warn people logits will not be returned anymore unless asked ie
+# os.environ['UNSLOTH_RETURN_LOGITS'] = '1'
+LOGITS_ERROR_STRING = \
+    "Unsloth: Logits are empty from 2024.11 onwards. To get raw logits again, please "\
+    'set the environment variable `UNSLOTH_RETURN_LOGITS` to `"1" BEFORE starting to train ie before `trainer.train()`. For example:\n'\
+    "```\nimport os\n"\
+    "os.environ['UNSLOTH_RETURN_LOGITS'] = '1'\n"\
+    "trainer.train()\n```\n"\
+    "No need to restart your console - just add `os.environ['UNSLOTH_RETURN_LOGITS'] = '1'` before trainer.train() and re-run the cell!"
+
+def raise_logits_error(*args, **kwargs): raise NotImplementedError(LOGITS_ERROR_STRING)
+def return_none(*args, **kwargs): return None
+class EmptyLogits:
+    def __init__(self): return
+    def raise_getattr_error(self, attr): return return_none if attr == "to" else raise_logits_error
+    __getitem__ = raise_logits_error
+    __getattr__ = raise_getattr_error
+    def __repr__(self): return LOGITS_ERROR_STRING
+    def __str__ (self): return LOGITS_ERROR_STRING
+pass
+EMPTY_LOGITS = EmptyLogits()
+functions = dir(torch.Tensor)
+for j, function in enumerate(functions):
+    if function.startswith("__") and function.endswith("__"):
+        exec(f"def raise_{j}(*args, **kwargs): print('{function}')", globals(), locals())
+        try: exec(f"EMPTY_LOGITS.{function} = raise_{j}", globals(), locals())
+        except: continue
+pass
+
+
+def mask_attention_mask_out(labels = None, attention_mask = None):
+    if labels is not None and attention_mask is not None:
+        attention_mask = attention_mask.to(device = labels.device)
+        labels[attention_mask == 0] = -100
+    return labels
+pass
+
+
+from torch import Tensor
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+from transformers.models.qwen3_moe.modeling_qwen3_moe import (__name__, F, Callable, Optional, Union, torch, nn, ACT2FN, Cache, GenerationMixin, use_kernel_forward_from_hub, FlashAttentionKwargs, MoeCausalLMOutputWithPast, MoeModelOutputWithPast, ROPE_INIT_FUNCTIONS, dynamic_rope_update, ALL_ATTENTION_FUNCTIONS, PreTrainedModel, Unpack, TransformersKwargs, can_return_tuple, deprecate_kwarg, Qwen3MoeConfig, Qwen3MoePreTrainedModel, Qwen3MoeModel, Qwen3MoeForCausalLM)
+
+@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
+def apply_rotary_pos_emb(q, k, cos, sin,  unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs: Unpack[TransformersKwargs],
+):
+    key_states = repeat_kv(key, module.num_key_value_groups)
+    value_states = repeat_kv(value, module.num_key_value_groups)
+
+    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+    if attention_mask is not None:
+        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype = torch.float32).to(attn_weights.dtype).to(query.dtype)
+    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
+    attn_output = torch.matmul(attn_weights, value_states)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+
+@torch.compiler.disable(recursive = False)
+@deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
+def Qwen3MoeAttention_forward(
+    self,
+    hidden_states: torch.Tensor,
+    position_embeddings: tuple[torch.Tensor, torch.Tensor],
+    attention_mask: Optional[torch.Tensor],
+    past_key_values: Optional[Cache] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    **kwargs: Unpack[FlashAttentionKwargs],
+) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
+    input_shape = hidden_states.shape[:-1]
+    hidden_shape = (*input_shape, -1, self.head_dim)
+
+    query_states = self.q_norm(self.q_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
+    key_states = self.k_norm(self.k_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
+    value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+
+    cos, sin = position_embeddings
+    query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+    if past_key_values is not None:
+        # sin and cos are specific to RoPE models; cache_position needed for the static cache
+        cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+        key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+    attention_interface: Callable = eager_attention_forward
+    if self.config._attn_implementation != "eager":
+        attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+
+    attn_output, attn_weights = attention_interface(
+        self,
+        query_states,
+        key_states,
+        value_states,
+        attention_mask,
+        dropout=0.0 if not self.training else self.attention_dropout,
+        scaling=self.scaling,
+        sliding_window=self.sliding_window,  # diff with Llama
+        **kwargs,
+    )
+
+    attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+    attn_output = self.o_proj(attn_output)
+    return attn_output, attn_weights
+
+class Qwen3MoeAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: Qwen3MoeConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
+        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = config.attention_dropout
+        self.is_causal = True
+
+        self.q_proj = nn.Linear(
+            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.k_proj = nn.Linear(
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.v_proj = nn.Linear(
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.o_proj = nn.Linear(
+            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
+        )
+        self.q_norm = Qwen3MoeRMSNorm(self.head_dim, eps=config.rms_norm_eps)  # unlike olmo, only on the head dim!
+        self.k_norm = Qwen3MoeRMSNorm(self.head_dim, eps=config.rms_norm_eps)  # thus post q_norm does not need reshape
+        self.sliding_window = getattr(config, "sliding_window", None)
+
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        past_key_values: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
+        return Qwen3MoeAttention_forward(self, hidden_states, position_embeddings, attention_mask, past_key_values, cache_position, **kwargs)
+
+
+@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
+def Qwen3MoeRMSNorm_forward(self, hidden_states):
+    input_dtype = hidden_states.dtype
+    hidden_states = hidden_states.to(torch.float32)
+    variance = hidden_states.pow(2).mean(-1, keepdim=True)
+    hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+    return self.weight * hidden_states.to(input_dtype)
+
+@use_kernel_forward_from_hub("RMSNorm")
+class Qwen3MoeRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        Qwen3MoeRMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        return Qwen3MoeRMSNorm_forward(self, hidden_states)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
+
+
+@torch.compile(fullgraph = False, dynamic = True, options = torch_compile_options)
+def Qwen3MoeMLP_forward(self, x):
+    down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+    return down_proj
+
+class Qwen3MoeMLP(nn.Module):
+    def __init__(self, config, intermediate_size=None):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = intermediate_size if intermediate_size is not None else config.intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        return Qwen3MoeMLP_forward(self, x)
+
+
+@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
+@torch.no_grad()
+@dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
+def Qwen3MoeRotaryEmbedding_forward(self, x, position_ids):
+    inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
+    position_ids_expanded = position_ids[:, None, :].float()
+
+    device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
+    with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+        freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+        emb = torch.cat((freqs, freqs), dim=-1)
+        cos = emb.cos() * self.attention_scaling
+        sin = emb.sin() * self.attention_scaling
+
+    return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+class Qwen3MoeRotaryEmbedding(nn.Module):
+    inv_freq: torch.Tensor  # fix linting for `register_buffer`
+
+    def __init__(self, config: Qwen3MoeConfig, device=None):
+        super().__init__()
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and isinstance(config.rope_scaling, dict):
+            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
+        else:
+            self.rope_type = "default"
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+
+    def forward(self, x, position_ids):
+        return Qwen3MoeRotaryEmbedding_forward(self, x, position_ids)
+
+
+@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
+def load_balancing_loss_func(
+    gate_logits: Union[torch.Tensor, tuple[torch.Tensor], None],
+    num_experts: Optional[int] = None,
+    top_k=2,
+    attention_mask: Optional[torch.Tensor] = None,
+) -> Union[torch.Tensor, int]:
+    r"""
+    Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.
+
+    See Switch Transformer (https://huggingface.co/papers/2101.03961) for more details. This function implements the loss
+    function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between
+    experts is too unbalanced.
+
+    Args:
+        gate_logits:
+            Logits from the `gate`, should be a tuple of model.config.num_hidden_layers tensors of
+            shape [batch_size X sequence_length, num_experts].
+        num_experts:
+            Number of experts
+        top_k:
+            The number of experts to route per-token, can be also interpreted as the `top-k` routing
+            parameter.
+        attention_mask (`torch.Tensor`, *optional*):
+            The attention_mask used in forward function
+            shape [batch_size X sequence_length] if not None.
+
+    Returns:
+        The auxiliary loss.
+    """
+    if gate_logits is None or not isinstance(gate_logits, tuple):
+        return 0
+
+    if isinstance(gate_logits, tuple):
+        compute_device = gate_logits[0].device
+        concatenated_gate_logits = torch.cat([layer_gate.to(compute_device) for layer_gate in gate_logits], dim=0)
+
+    routing_weights = torch.nn.functional.softmax(concatenated_gate_logits, dim=-1, dtype = torch.float32).to(concatenated_gate_logits.dtype)
+
+    _, selected_experts = torch.topk(routing_weights, top_k, dim=-1)
+
+    expert_mask = torch.nn.functional.one_hot(selected_experts, num_experts)
+
+    if attention_mask is None:
+        # Compute the percentage of tokens routed to each experts
+        tokens_per_expert = torch.mean(expert_mask.float(), dim=0)
+
+        # Compute the average probability of routing to these experts
+        router_prob_per_expert = torch.mean(routing_weights, dim=0)
+    else:
+        batch_size, sequence_length = attention_mask.shape
+        num_hidden_layers = concatenated_gate_logits.shape[0] // (batch_size * sequence_length)
+
+        # Compute the mask that masks all padding tokens as 0 with the same shape of expert_mask
+        expert_attention_mask = (
+            attention_mask[None, :, :, None, None]
+            .expand((num_hidden_layers, batch_size, sequence_length, top_k, num_experts))
+            .reshape(-1, top_k, num_experts)
+            .to(compute_device)
+        )
+
+        # Compute the percentage of tokens routed to each experts
+        tokens_per_expert = torch.sum(expert_mask.float() * expert_attention_mask, dim=0) / torch.sum(
+            expert_attention_mask, dim=0
+        )
+
+        # Compute the mask that masks all padding tokens as 0 with the same shape of tokens_per_expert
+        router_per_expert_attention_mask = (
+            attention_mask[None, :, :, None]
+            .expand((num_hidden_layers, batch_size, sequence_length, num_experts))
+            .reshape(-1, num_experts)
+            .to(compute_device)
+        )
+
+        # Compute the average probability of routing to these experts
+        router_prob_per_expert = torch.sum(routing_weights * router_per_expert_attention_mask, dim=0) / torch.sum(
+            router_per_expert_attention_mask, dim=0
+        )
+
+    overall_loss = torch.sum(tokens_per_expert * router_prob_per_expert.unsqueeze(0))
+    return overall_loss * num_experts
+
+
+@torch.compiler.disable(recursive = False)
+@can_return_tuple
+def Qwen3MoeForCausalLM_forward(
+    self,
+    input_ids: Optional[torch.LongTensor] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[Cache] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_router_logits: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    **kwargs: Unpack[TransformersKwargs],
+) -> MoeCausalLMOutputWithPast:
+    r"""
+    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]`.
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, Qwen3MoeForCausalLM
+
+    >>> model = Qwen3MoeForCausalLM.from_pretrained("Qwen/Qwen3-MoE-15B-A2B")
+    >>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-MoE-15B-A2B")
+
+    >>> 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_router_logits = (
+        output_router_logits if output_router_logits is not None else self.config.output_router_logits
+    )
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs: MoeModelOutputWithPast = 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_router_logits=output_router_logits,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    hidden_states = outputs.last_hidden_state
+    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+    slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+    logits = self.lm_head(hidden_states[:, slice_indices, :]) if os.environ.get('UNSLOTH_RETURN_LOGITS', '0') == '1' else EMPTY_LOGITS
+    loss = None
+    NOT_RETURN_LOGITS = os.environ.get('UNSLOTH_RETURN_LOGITS', '0') == '0'
+    RETURN_HIDDEN_STATES = os.environ.get("UNSLOTH_RETURN_HIDDEN_STATES", "0") == "1"
+    
+    n_items = None
+    if (kwargs) != () and type(kwargs) is dict:
+        n_items = (kwargs).get("num_items_in_batch", None) or (kwargs).get("n_items", None)
+    if n_items is None:
+        all_locals = locals()
+        if 'loss_kwargs' in all_locals:
+            __kwargs = all_locals['loss_kwargs']
+            if type(__kwargs) is dict:
+                n_items = __kwargs.get("num_items_in_batch", None)
+                if n_items is None: n_items = __kwargs.get("n_items", None)
+        if n_items is None and 'kwargs' in all_locals:
+            __kwargs = all_locals['kwargs']
+            if type(__kwargs) is dict:
+                n_items = __kwargs.get("num_items_in_batch", None)
+                if n_items is None: n_items = __kwargs.get("n_items", None)
+        if n_items is None:
+            all_locals = all_locals.values()
+            for __kwargs in all_locals:
+                if type(__kwargs) is dict:
+                    n_items = __kwargs.get("num_items_in_batch", None)
+                    if n_items is None: n_items = __kwargs.get("n_items", None)
+                    break
+    pass
+    
+    requires_grad_ = self.lm_head.weight.requires_grad
+    requires_grad_ = requires_grad_ or self.lm_head.weight.dtype == torch.float32
+    
+    if RETURN_HIDDEN_STATES:
+        logits = hidden_states[:, slice_indices, :]
+    elif labels is None:
+        
+    
+        # Set compiler stance to fail on recompiles for inference
+        global INFERENCE_RUNS
+        if torch_dynamo_eval_frame is not None:
+            old_stance = torch_dynamo_eval_frame._stance.stance
+        else:
+            old_stance = None
+        if old_stance is not None and INFERENCE_RUNS == 1:
+            # Skip guards and return to eager -> we still need guards!
+            torch_compiler_set_stance(stance = "eager_on_recompile", skip_guard_eval_unsafe = False)
+            if UNSLOTH_ENABLE_LOGGING:
+                logger_compiler.info(
+                    f"Unsloth: Removing compiler guards after 1 inference run. "\
+                    f"DYNAMO_STANCE.stance = {torch_dynamo_eval_frame._stance.stance} "\
+                    f"DYNAMO_STANCE.skip_guard_eval_unsafe = {torch_dynamo_eval_frame._stance.skip_guard_eval_unsafe}"
+                )
+        elif old_stance == "eager_on_recompile":
+            pass
+        elif old_stance == "default" and INFERENCE_RUNS > 1:
+            # Reset compiler stance
+            torch_compiler_set_stance(stance = "default", skip_guard_eval_unsafe = False)
+            if UNSLOTH_ENABLE_LOGGING:
+                logger_compiler.info(
+                    f"Unsloth: Reseting guards. "\
+                    f"DYNAMO_STANCE.stance = {torch_dynamo_eval_frame._stance.stance} "\
+                    f"DYNAMO_STANCE.skip_guard_eval_unsafe = {torch_dynamo_eval_frame._stance.skip_guard_eval_unsafe}"
+                )
+            INFERENCE_RUNS = 0
+        INFERENCE_RUNS += 1
+    
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+    elif (() == () and () == ()) and (UNSLOTH_ENABLE_CCE) and NOT_RETURN_LOGITS and self.loss_function.__name__.endswith("ForCausalLMLoss") and labels is not None and not requires_grad_:
+        loss = fused_linear_cross_entropy(
+            hidden_states      = hidden_states[:, slice_indices, :],
+            lm_weight          = self.lm_head.weight,
+            labels             = labels.to(self.lm_head.weight.device),
+            num_items_in_batch = n_items,
+            logit_softcapping  = None if () == () else (),
+        )
+    elif self.loss_function.__name__.endswith("ForCausalLMLoss") and labels is not None:
+        lm_head_weight = self.lm_head.weight
+        lm_head_bias   = getattr(self.lm_head, "bias", None)
+    
+        # ========= NEW fused =========
+        _hidden_states = hidden_states[:, slice_indices, :]
+        torch._dynamo.mark_dynamic(_hidden_states, 1)
+        torch._dynamo.mark_dynamic(labels, 1)
+        loss = unsloth_fused_ce_loss(
+            trainer              = None,
+            hidden_states        = _hidden_states,
+            lm_head_weight       = lm_head_weight,
+            lm_head_bias         = lm_head_bias,
+            labels               = labels,
+            mask                 = None,
+            n_items              = n_items,
+            scaling              = getattr(self, "accelerator_scaler", None),
+            target_gb            = None,
+            torch_compile        = not UNSLOTH_COMPILE_DISABLE,
+            logit_scale_multiply = () if () != () else 0,
+            logit_scale_divide   = () if () != () else 0,
+            logit_softcapping    = () if () != () else 0,
+        )
+    else:
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        if () != ():
+            logits = logits * ()
+        if () != ():
+            logits = logits / ()
+        if () not in (None, (),):
+            logits = logits / ()
+            logits = torch.tanh(logits)
+            logits = logits * ()
+        loss = self.loss_function(logits, labels.to(self.lm_head.weight.device), vocab_size=self.vocab_size, **kwargs)
+
+
+    aux_loss = None
+    if output_router_logits:
+        aux_loss = load_balancing_loss_func(
+            outputs.router_logits,
+            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
+
+    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,
+    )
+
+class Qwen3MoeForCausalLM(Qwen3MoePreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+    _tp_plan = {"lm_head": "colwise_rep"}
+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = Qwen3MoeModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.router_aux_loss_coef = config.router_aux_loss_coef
+        self.num_experts = config.num_experts
+        self.num_experts_per_tok = config.num_experts_per_tok
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_router_logits: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> MoeCausalLMOutputWithPast:
+        return Qwen3MoeForCausalLM_forward(self, input_ids, attention_mask, position_ids, past_key_values, inputs_embeds, labels, use_cache, output_router_logits, cache_position, logits_to_keep, **kwargs)