liger-kernel-nightly 0.6.1.dev20250805235815__py3-none-any.whl → 0.6.1.dev20250809233744__py3-none-any.whl

liger_kernel/ops/llama4_rope.py

@@ -0,0 +1,225 @@
+import torch
+import triton
+import triton.language as tl
+
+
+def _prepare_freqs(freqs_cis: torch.Tensor, seq_len: int, head_dim_half: int):
+    # Split or unpack complex frequencies into real and imag parts
+    if freqs_cis.is_complex():
+        freqs_real = freqs_cis.real
+        freqs_imag = freqs_cis.imag
+    else:
+        # Already split: last dim should be 2*head_dim_half
+        if freqs_cis.shape[-1] == 2 * head_dim_half:
+            freqs_real = freqs_cis[..., :head_dim_half]
+            freqs_imag = freqs_cis[..., head_dim_half:]
+        else:
+            raise ValueError(
+                f"Unexpected freqs_cis shape for non-complex input: {freqs_cis.shape}, expected last dim = {2 * head_dim_half}"
+            )
+
+    # Canonicalize to shape (seq_len, head_dim_half):
+    # 1) Ensure the last dimension is head_dim_half
+    if freqs_real.shape[-1] != head_dim_half:
+        raise ValueError(f"Unexpected last dim for freqs: {freqs_real.shape[-1]} (expected {head_dim_half})")
+    # 2) Flatten all leading dims to a single row dimension
+    freqs_real = freqs_real.reshape(-1, head_dim_half)
+    freqs_imag = freqs_imag.reshape(-1, head_dim_half)
+    # 3) If we have fewer rows than seq_len, allow broadcasting when single row
+    if freqs_real.shape[0] < seq_len:
+        if freqs_real.shape[0] == 1:
+            freqs_real = freqs_real.expand(seq_len, -1)
+            freqs_imag = freqs_imag.expand(seq_len, -1)
+        else:
+            raise ValueError(f"Insufficient rows in freqs: {freqs_real.shape[0]} < seq_len={seq_len}")
+    # 4) If we have more rows than seq_len (e.g., batch present), take the first seq_len rows
+    elif freqs_real.shape[0] > seq_len:
+        freqs_real = freqs_real[:seq_len]
+        freqs_imag = freqs_imag[:seq_len]
+
+    return freqs_real, freqs_imag
+
+
+def _maybe_to_dtype(t: torch.Tensor, dtype: torch.dtype) -> torch.Tensor:
+    return t if t.dtype == dtype else t.to(dtype)
+
+
+def _maybe_contiguous(t: torch.Tensor) -> torch.Tensor:
+    return t if t.is_contiguous() else t.contiguous()
+
+
+def _cast_and_contiguous(q, k, freqs_real, freqs_imag):
+    # Choose compute dtype: use fp32 only when inputs are fp32; otherwise keep input dtype for performance
+    compute_dtype = torch.float32 if q.dtype == torch.float32 else q.dtype
+
+    # Make sure q/k share the same dtype before casting to compute dtype
+    if k.dtype != q.dtype:
+        k = k.to(q.dtype)
+
+    q = _maybe_contiguous(_maybe_to_dtype(q, compute_dtype))
+    k = _maybe_contiguous(_maybe_to_dtype(k, compute_dtype))
+    freqs_real = _maybe_contiguous(_maybe_to_dtype(freqs_real, compute_dtype))
+    freqs_imag = _maybe_contiguous(_maybe_to_dtype(freqs_imag, compute_dtype))
+    return q, k, freqs_real, freqs_imag
+
+
+@triton.jit
+def _llama4_rope_kernel(
+    q_ptr,
+    k_ptr,
+    freqs_real_ptr,
+    freqs_imag_ptr,
+    q_row_stride,
+    k_row_stride,
+    q_head_stride,
+    k_head_stride,
+    freqs_row_stride,
+    seq_len,
+    batch_size,
+    imag_sign,
+    head_dim_half: tl.constexpr,
+    n_q_heads: tl.constexpr,
+    n_k_heads: tl.constexpr,
+    BLOCK_SIZE: tl.constexpr,
+):
+    """
+    H100-optimized RoPE kernel with improved parallelization across heads and dimensions.
+    Grid: (batch*seq, head)
+    """
+    # 2D grid
+    pid_bs = tl.program_id(0)  # over batch*seq
+    pid_h = tl.program_id(1)  # over heads
+
+    batch_idx = pid_bs // seq_len
+    seq_idx = pid_bs % seq_len
+
+    # Bounds check
+    if batch_idx >= batch_size or seq_idx >= seq_len:
+        return
+
+    # Base pointers for this (batch, seq) position
+    base_offset = batch_idx * seq_len + seq_idx
+    q_base = q_ptr + base_offset * q_row_stride
+    k_base = k_ptr + base_offset * k_row_stride
+
+    # Tiling over dim/2
+    for d_start in tl.static_range(0, head_dim_half, BLOCK_SIZE):
+        d_indices = d_start + tl.arange(0, BLOCK_SIZE)
+        mask_d = d_indices < head_dim_half
+
+        # Load frequencies once per tile (freqs layout: [seq_len, head_dim_half])
+        freq_idx = d_indices
+        freqs_real = tl.load(freqs_real_ptr + seq_idx * freqs_row_stride + freq_idx, mask=mask_d, other=0.0)
+        freqs_imag = tl.load(freqs_imag_ptr + seq_idx * freqs_row_stride + freq_idx, mask=mask_d, other=0.0)
+        freqs_imag = freqs_imag * imag_sign
+
+        # Process one query head per program in pid_h
+        if pid_h < n_q_heads:
+            q_head_ptr = q_base + pid_h * q_head_stride
+            q_real = tl.load(q_head_ptr + d_indices * 2, mask=mask_d, other=0.0)
+            q_imag = tl.load(q_head_ptr + d_indices * 2 + 1, mask=mask_d, other=0.0)
+
+            # Complex multiply with FMAs: (a+ib)*(c+i d) = (a*c - b*d) + i(a*d + b*c)
+            new_q_real = tl.math.fma(q_real, freqs_real, -(q_imag * freqs_imag))
+            new_q_imag = tl.math.fma(q_real, freqs_imag, q_imag * freqs_real)
+
+            tl.store(q_head_ptr + d_indices * 2, new_q_real, mask=mask_d)
+            tl.store(q_head_ptr + d_indices * 2 + 1, new_q_imag, mask=mask_d)
+
+        # Process one key head per program in pid_h
+        if pid_h < n_k_heads:
+            k_head_ptr = k_base + pid_h * k_head_stride
+            k_real = tl.load(k_head_ptr + d_indices * 2, mask=mask_d, other=0.0)
+            k_imag = tl.load(k_head_ptr + d_indices * 2 + 1, mask=mask_d, other=0.0)
+
+            new_k_real = tl.math.fma(k_real, freqs_real, -(k_imag * freqs_imag))
+            new_k_imag = tl.math.fma(k_real, freqs_imag, k_imag * freqs_real)
+
+            tl.store(k_head_ptr + d_indices * 2, new_k_real, mask=mask_d)
+            tl.store(k_head_ptr + d_indices * 2 + 1, new_k_imag, mask=mask_d)
+
+
+def _select_kernel_meta(head_dim_half: int):
+    # Heuristic tuning for block size and num_warps
+    if head_dim_half >= 256:
+        return 128, 8
+    if head_dim_half >= 96:
+        return 128, 4
+    if head_dim_half >= 48:
+        return 64, 4
+    if head_dim_half >= 24:
+        return 32, 2
+    return 16, 2
+
+
+def llama4_rope_forward(q, k, freqs_cis, BLOCK_SIZE: int = None, imag_sign: float = 1.0):
+    # Save original dtype for casting back
+    original_dtype = q.dtype
+
+    batch_size, seq_len, n_q_heads, head_dim = q.shape
+    _, _, n_k_heads, _ = k.shape
+    head_dim_half = head_dim // 2
+
+    # Prepare frequencies
+    freqs_real, freqs_imag = _prepare_freqs(freqs_cis, seq_len, head_dim_half)
+
+    # Cast to appropriate dtype and make contiguous only when needed
+    q, k, freqs_real, freqs_imag = _cast_and_contiguous(q, k, freqs_real, freqs_imag)
+
+    # H100-optimized meta-params
+    if BLOCK_SIZE is None:
+        BLOCK_SIZE, num_warps = _select_kernel_meta(head_dim_half)
+    else:
+        # Provide a default num_warps if caller pins BLOCK_SIZE
+        _, num_warps = _select_kernel_meta(head_dim_half)
+
+    # 2D grid: one program per (batch, seq, head)
+    n_heads_max = max(n_q_heads, n_k_heads)
+    grid = (batch_size * seq_len, n_heads_max)
+
+    # Launch kernel
+    _llama4_rope_kernel[grid](
+        q,
+        k,
+        freqs_real,
+        freqs_imag,
+        q.stride(1),
+        k.stride(1),
+        q.stride(2),
+        k.stride(2),
+        freqs_real.stride(0),
+        seq_len,
+        batch_size,
+        imag_sign,
+        head_dim_half,
+        n_q_heads,
+        n_k_heads,
+        BLOCK_SIZE,
+        num_warps=num_warps,
+        num_stages=2,
+    )
+
+    # Cast back to original dtype only if it differs from compute dtype
+    if q.dtype != original_dtype:
+        q = q.to(original_dtype)
+    if k.dtype != original_dtype:
+        k = k.to(original_dtype)
+
+    return q, k
+
+
+class LigerLlama4RopeFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, q, k, freqs_cis, BLOCK_SIZE: int = None):
+        q_out, k_out = llama4_rope_forward(q, k, freqs_cis, BLOCK_SIZE, imag_sign=1.0)
+        ctx.save_for_backward(freqs_cis.detach() if isinstance(freqs_cis, torch.Tensor) else freqs_cis)
+        ctx.BLOCK_SIZE = BLOCK_SIZE
+        return q_out, k_out
+
+    @staticmethod
+    def backward(ctx, dq, dk):
+        (freqs_cis,) = ctx.saved_tensors
+        BLOCK_SIZE = getattr(ctx, "BLOCK_SIZE", None)
+        # Use imag_sign=-1.0 for conjugate without materializing a new tensor
+        dq_out, dk_out = llama4_rope_forward(dq, dk, freqs_cis, BLOCK_SIZE, imag_sign=-1.0)
+        return dq_out, dk_out, None

liger_kernel/transformers/__init__.py

@@ -11,6 +11,8 @@ from liger_kernel.transformers.fused_linear_jsd import LigerFusedLinearJSD  # no
 from liger_kernel.transformers.geglu import LigerGEGLUMLP  # noqa: F401
 from liger_kernel.transformers.jsd import LigerJSD  # noqa: F401
 from liger_kernel.transformers.layer_norm import LigerLayerNorm  # noqa: F401
+from liger_kernel.transformers.llama4_rope import liger_llama4_text_rotary_pos_emb  # noqa: F401
+from liger_kernel.transformers.llama4_rope import liger_llama4_vision_rotary_pos_emb  # noqa: F401
 from liger_kernel.transformers.rms_norm import LigerRMSNorm  # noqa: F401
 from liger_kernel.transformers.rope import liger_rotary_pos_emb  # noqa: F401
 from liger_kernel.transformers.swiglu import LigerBlockSparseTop2MLP  # noqa: F401
@@ -125,6 +127,8 @@ __all__ = [
     "LigerFusedAddRMSNorm",
     "LigerRMSNorm",
     "liger_rotary_pos_emb",
+    "liger_llama4_text_rotary_pos_emb",
+    "liger_llama4_vision_rotary_pos_emb",
     "LigerBlockSparseTop2MLP",
     "LigerPhi3SwiGLUMLP",
     "LigerQwen3MoeSwiGLUMLP",

liger_kernel/transformers/llama4_rope.py

@@ -0,0 +1,93 @@
+"""
+Liger Kernel implementation of Llama4 Rotary Position Embedding (RoPE).
+Supports both text and vision RoPE variants with fused operations for optimal performance.
+"""
+
+import torch
+
+from liger_kernel.ops.llama4_rope import LigerLlama4RopeFunction
+
+
+def liger_llama4_text_rotary_pos_emb(
+    xq: torch.Tensor,
+    xk: torch.Tensor,
+    freqs_cis: torch.Tensor,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """
+    Liger-optimized implementation of Llama4 text rotary position embedding.
+
+    This implementation uses a fused Triton kernel for complex multiplication,
+    providing significant performance improvements over the original PyTorch implementation.
+
+    Args:
+        xq (torch.Tensor): Query tensor of shape (batch_size, seq_len, num_heads, head_dim)
+        xk (torch.Tensor): Key tensor of shape (batch_size, seq_len, num_heads, head_dim)
+        freqs_cis (torch.Tensor): Complex frequency tensor from Llama4TextRotaryEmbedding
+
+    Returns:
+        Tuple[torch.Tensor, torch.Tensor]: Rotated query and key tensors
+    """
+    # Use fused Triton kernel for complex RoPE
+    return LigerLlama4RopeFunction.apply(xq, xk, freqs_cis)
+
+
+def liger_llama4_vision_rotary_pos_emb(
+    query: torch.Tensor,
+    key: torch.Tensor,
+    freqs_ci: torch.Tensor,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """
+    Liger-optimized implementation of Llama4 vision rotary position embedding.
+
+    This implementation uses the same fused Triton kernel as text RoPE,
+    providing performance improvements for vision transformer attention.
+
+    Args:
+        query (torch.Tensor): Query tensor of shape (batch_size, seq_len, num_heads, head_dim)
+        key (torch.Tensor): Key tensor of shape (batch_size, seq_len, num_heads, head_dim)
+        freqs_ci (torch.Tensor): Complex frequency tensor for 2D positions
+
+    Returns:
+        Tuple[torch.Tensor, torch.Tensor]: Rotated query and key tensors
+    """
+    # Handle broadcasting for vision RoPE
+    if freqs_ci.dim() == 3:
+        try:
+            # Try the regular 3D expansion
+            freqs_ci = freqs_ci.unsqueeze(0).expand(query.shape[0], -1, -1)
+        except RuntimeError as e:
+            if "expand" in str(e) and "4" in str(e):
+                # The tensor is actually 4D internally, handle it differently
+                freqs_ci = freqs_ci.squeeze(1)  # Remove the middle dimension
+                freqs_ci = freqs_ci.unsqueeze(0).expand(query.shape[0], -1, -1)
+            else:
+                raise e
+    elif freqs_ci.dim() == 4:  # (1, seq_len, 1, head_dim//2) - already properly shaped
+        # Squeeze the middle dimension to get (1, seq_len, head_dim//2)
+        freqs_ci = freqs_ci.squeeze(2)
+    elif freqs_ci.dim() == 2:  # (seq_len, head_dim//2) - needs expansion
+        freqs_ci = freqs_ci.unsqueeze(0).expand(query.shape[0], -1, -1)
+    else:
+        raise ValueError(f"Unexpected freqs_ci shape: {freqs_ci.shape}")
+
+    # Use the same fused kernel as text RoPE
+    return LigerLlama4RopeFunction.apply(query, key, freqs_ci)
+
+
+# Note: We only patch the functions, not the classes
+# The original Llama4TextRotaryEmbedding and Llama4VisionRotaryEmbedding classes remain unchanged
+
+
+# Convenience functions for monkey patching
+def apply_liger_llama4_rope_full(modeling_module):
+    """
+    Apply Liger optimizations to Llama4 RoPE functions.
+
+    Args:
+        modeling_module: The transformers modeling module to patch
+    """
+    # Replace the text RoPE function
+    modeling_module.apply_rotary_emb = liger_llama4_text_rotary_pos_emb
+
+    # Replace the vision RoPE function
+    modeling_module.vision_apply_rotary_emb = liger_llama4_vision_rotary_pos_emb

liger_kernel/transformers/model/phi3.py

@@ -5,131 +5,12 @@ from typing import Union
 
 import torch
 
-from torch.nn import CrossEntropyLoss
+from transformers.modeling_outputs import BaseModelOutputWithPast
 from transformers.modeling_outputs import CausalLMOutputWithPast
-from transformers.utils.deprecation import deprecate_kwarg
 
-from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
 from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
 
 
-def lce_forward_deprecated(
-    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,
-    return_dict: Optional[bool] = None,
-    cache_position: Optional[torch.LongTensor] = None,
-    skip_logits: Optional[bool] = None,
-) -> Union[Tuple, CausalLMOutputWithPast]:
-    r"""
-    Copy paste phi3 forward from transfomers v4.44.2 but replace torch cross entropy with liger fused linear cross entropy
-
-
-    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]`.
-
-    Returns:
-
-    Example:
-
-    ```python
-    >>> from transformers import AutoTokenizer, Phi3ForCausalLM
-
-    >>> model = Phi3ForCausalLM.from_pretrained("microsoft/phi-3-mini-4k-instruct")
-    >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-3-mini-4k-instruct")
-
-    >>> prompt = "This is an example script ."
-    >>> 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]
-    'This is an example script .\n Certainly! Below is a sample script that demonstrates a simple task, such as calculating the sum'
-    ```"""
-
-    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,
-    )
-
-    hidden_states = outputs[0]
-
-    loss = None
-    logits = None
-
-    if skip_logits and labels is None:
-        raise ValueError("skip_logits is True, but labels is None")
-
-    if skip_logits is None:
-        # By default, if in training mode, don't materialize logits
-        skip_logits = self.training and labels is not None
-
-    if skip_logits:
-        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:
-        logits = self.lm_head(hidden_states)
-
-        loss = None
-        if labels is not None:
-            # Upcast to float if we need to compute the loss to avoid potential precision issues
-            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)
-
-    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,
-    )
-
-
-@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,
@@ -148,36 +29,21 @@ def lce_forward(
     **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]`.
-
-        logits_to_keep (`int` or `torch.Tensor`, *optional*):
-            If an `int`, compute logits for the last `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.
-            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
-            This is useful when using packed tensor format (single dimension for batch and sequence length).
-
-    Returns:
-
     Example:
 
     ```python
     >>> from transformers import AutoTokenizer, Phi3ForCausalLM
 
-    >>> model = Phi3ForCausalLM.from_pretrained("microsoft/phi-3-mini-4k-instruct")
-    >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-3-mini-4k-instruct")
+    >>> model = Phi3ForCausalLM.from_pretrained("meta-phi3/Phi3-2-7b-hf")
+    >>> tokenizer = AutoTokenizer.from_pretrained("meta-phi3/Phi3-2-7b-hf")
 
-    >>> prompt = "This is an example script ."
+    >>> 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]
-    'This is an example script .\n Certainly! Below is a sample script that demonstrates a simple task, such as calculating the sum'
+    "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
@@ -186,21 +52,18 @@ def lce_forward(
     )
     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(
+    outputs: BaseModelOutputWithPast = 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,
         **kwargs,
     )
 
-    hidden_states = outputs[0]
+    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
     kept_hidden_states = hidden_states[:, slice_indices, :]

liger_kernel/transformers/monkey_patch.py

@@ -26,7 +26,6 @@ from liger_kernel.transformers.model.mistral import lce_forward as mistral_lce_f
 from liger_kernel.transformers.model.mixtral import lce_forward as mixtral_lce_forward
 from liger_kernel.transformers.model.mixtral import lce_forward_deprecated as mixtral_lce_forward_deprecated
 from liger_kernel.transformers.model.phi3 import lce_forward as phi3_lce_forward
-from liger_kernel.transformers.model.phi3 import lce_forward_deprecated as phi3_lce_forward_deprecated
 from liger_kernel.transformers.model.qwen2 import lce_forward as qwen2_lce_forward
 from liger_kernel.transformers.model.qwen2 import lce_forward_deprecated as qwen2_lce_forward_deprecated
 from liger_kernel.transformers.model.smollm3 import lce_forward as smollm3_lce_forward
@@ -449,7 +448,7 @@ def apply_liger_kernel_to_llava(
 
 
 def apply_liger_kernel_to_llama4(
-    rope: bool = False,
+    rope: bool = True,
     cross_entropy: bool = False,
     fused_linear_cross_entropy: bool = True,
     rms_norm: bool = True,
@@ -485,7 +484,9 @@ def apply_liger_kernel_to_llama4(
     from liger_kernel.transformers.model.llama4 import lce_forward as llama4_lce_forward
 
     if rope:
-        raise NotImplementedError("liger_rotary_pos_emb is not available for Llama4 models.")
+        from liger_kernel.transformers.llama4_rope import apply_liger_llama4_rope_full
+
+        apply_liger_llama4_rope_full(modeling_llama4)
     if rms_norm:
         modeling_llama4.Llama4TextRMSNorm = LigerRMSNorm
     if swiglu:
@@ -1675,25 +1676,14 @@ def apply_liger_kernel_to_phi3(
     if swiglu:
         modeling_phi3.Phi3MLP = LigerPhi3SwiGLUMLP
     if cross_entropy:
-        if transformer_version >= version.parse(SUPPORTED_TRANSFORMER_VERSION):
-            from transformers.loss.loss_utils import nn
+        from transformers.loss.loss_utils import nn
 
-            nn.functional.cross_entropy = liger_cross_entropy
-        else:
-            logger.warning(TRANSFORMER_DEPRECATION_WARNING)
-            modeling_phi3.CrossEntropyLoss = LigerCrossEntropyLoss
+        nn.functional.cross_entropy = liger_cross_entropy
     if fused_linear_cross_entropy:
-        if transformer_version >= version.parse(SUPPORTED_TRANSFORMER_VERSION):
-            if model is not None:
-                model.forward = MethodType(phi3_lce_forward, model)
-            else:
-                modeling_phi3.Phi3ForCausalLM.forward = phi3_lce_forward
-        else:  # if version < 4.46.1
-            logger.warning(TRANSFORMER_DEPRECATION_WARNING)
-            if model is not None:
-                model.forward = MethodType(phi3_lce_forward_deprecated, model)
-            else:
-                modeling_phi3.Phi3ForCausalLM.forward = phi3_lce_forward_deprecated
+        if model is not None:
+            model.forward = MethodType(phi3_lce_forward, model)
+        else:
+            modeling_phi3.Phi3ForCausalLM.forward = phi3_lce_forward
 
     if model is not None:
         # The model instance already exists, so we need to additionally patch the

{liger_kernel_nightly-0.6.1.dev20250805235815.dist-info → liger_kernel_nightly-0.6.1.dev20250809233744.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.1.dev20250805235815
+Version: 0.6.1.dev20250809233744
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.6.1.dev20250805235815.dist-info → liger_kernel_nightly-0.6.1.dev20250809233744.dist-info}/RECORD

@@ -29,6 +29,7 @@ liger_kernel/ops/grpo_loss.py,sha256=anRnv7k1-AV3pCC6_TqP0GMg78YYUfRAJrbpx6PVhl0
 liger_kernel/ops/jsd.py,sha256=onHp5T3MbvJaVz5Vup7Ww6EQp_HTaZeayTjJk6FgQMY,7042
 liger_kernel/ops/kl_div.py,sha256=ZjGdDLKWksHT9dZ0xF_TDgAkj5cuMTwwT5tr9E-_24o,8734
 liger_kernel/ops/layer_norm.py,sha256=BHPDuaogMTfIJkBJdqLZbOQouNWTf3fJVyOQOD7blCE,9901
+liger_kernel/ops/llama4_rope.py,sha256=-aqdZzllklTN8b9--e-TsWY_ntGCN8-tyseT4x0bd8s,8223
 liger_kernel/ops/multi_token_attention.py,sha256=Oz_RXDp-OSS_R_HuGmaETHdAJ7Toda_70OfE7TXMUlY,7645
 liger_kernel/ops/qwen2vl_mrope.py,sha256=3GExhYpLgB4VUtyZyjRk8XjEur3W4EWF6HQ67ML5vBU,8481
 liger_kernel/ops/rms_norm.py,sha256=DtvsWN5YktFAoc0JYSAwVeoZfryBFJlX-ipU7ooP01A,18891
@@ -40,7 +41,7 @@ liger_kernel/ops/tvd.py,sha256=FHJtLQI95ijqgg9UtaHpMAjSCiPxB6CduPwPMcGxelc,6405
 liger_kernel/ops/utils.py,sha256=uoFKQqo-34N2TWQNvXMFywqGiOMMXNEVBxVojzlUAa0,3836
 liger_kernel/ops/experimental/embedding.py,sha256=tolj3tItkzpSb30zWqDN2_yX4ectflaQ8HMyKyFIQc8,4172
 liger_kernel/ops/experimental/mm_int8int2.py,sha256=TrS9lpwekrik_w5qE7AhMJD1bcq-OidjtbsW80oZ6IM,13314
-liger_kernel/transformers/__init__.py,sha256=VoHQp5emsAJAouql37RuvtGFeZCoMIHgoIxfsyYMTc8,7564
+liger_kernel/transformers/__init__.py,sha256=YQ3ffAaZWLy266snmFFHHfoz4EX1AhcSfojXZhOs6h0,7842
 liger_kernel/transformers/auto_model.py,sha256=0qCTRZt280Bj_LcFdzo9hlaR-BWNazawXOGgoCZjgEg,1545
 liger_kernel/transformers/cross_entropy.py,sha256=z3KTWQnFxr_IZaVjtYt0ZNEWQdDdYThN35xWkHlDGH0,1683
 liger_kernel/transformers/dyt.py,sha256=i-4GPaMrl-jab9TVI5qN0-H9qycn_mCbV82ozU4nbmU,723
@@ -56,7 +57,8 @@ liger_kernel/transformers/grpo_loss.py,sha256=uAkUNKSnUGEOqa82L9w2e6AI1kcmG8K45-
 liger_kernel/transformers/jsd.py,sha256=DGqRnxIZxsvxo0_tbbxX3b-sDbDjC_yKufyRIHCcScY,2979
 liger_kernel/transformers/kl_div.py,sha256=WLffFbh1EExD2Eb1F7lN11fo9JJC-0751WJjZAF1Fj8,409
 liger_kernel/transformers/layer_norm.py,sha256=c9pk3PEasOKYR0rhe5e5nNrnYKVCEW4VC8S6LpCq9EQ,906
-liger_kernel/transformers/monkey_patch.py,sha256=tXKo4EKVp3szpdqPh051oLZFrlg_hCbWRv0RpSX_kfY,89238
+liger_kernel/transformers/llama4_rope.py,sha256=kS6PSHEwf3dS7hD7C7p8S0geugx2EMCiP0h0F7LsUoY,3639
+liger_kernel/transformers/monkey_patch.py,sha256=zeYj1XhI1at_gWdKnvZqazT53uBy_YOV36ZuQfnhf20,88545
 liger_kernel/transformers/multi_token_attention.py,sha256=l9VDICK0dfmifUDW668hGscP8AHq2rYcM2oGUa3baRQ,1751
 liger_kernel/transformers/qwen2vl_mrope.py,sha256=5EwSqrMdsL9MYspeBMXBsNJKvH0MOmRrtJXAJlnnlOI,1047
 liger_kernel/transformers/rms_norm.py,sha256=vkekcvTeWY8vL4H6hg3t0XeY0Ew_3OFMPHuzqlxPPVw,2719
@@ -81,7 +83,7 @@ liger_kernel/transformers/model/mixtral.py,sha256=VY-y73IyjcCyWyI7ahxXLw0fJrhgjY
 liger_kernel/transformers/model/mllama.py,sha256=my29NXk-p6ckQaP8qDIN8e318yI_9mQZHt38MV3SqLY,11280
 liger_kernel/transformers/model/olmo2.py,sha256=6L_bo-ZUgO1lYppdJneOtYxNIylQKS6BiGp13g7Uq9E,5259
 liger_kernel/transformers/model/paligemma.py,sha256=xuIx3oOwTgftU3jqLfWOxUxgCLBNJh0yNC21an9qDjo,18773
-liger_kernel/transformers/model/phi3.py,sha256=aOl1Pz2rp5jSahRKUHKFPgkdkgG28fnHPpOW2ZVnMPg,10124
+liger_kernel/transformers/model/phi3.py,sha256=AwScxUe3LjmHHyQg4gW9bMoUI7uA6fUEMXJ3YhBiHtQ,4046
 liger_kernel/transformers/model/qwen2.py,sha256=3fpOTEOkniQmkCfN1KUa3KhseHJVzhj2Ht9FdYPUy-E,9962
 liger_kernel/transformers/model/qwen2_5_vl.py,sha256=zEVVwotCXnAm3RRc8-1Nc8uitSWrwW4B9dYY2uOZDwg,6331
 liger_kernel/transformers/model/qwen2_vl.py,sha256=5vK-vtCDpKZ2w33xYp2BS8kQYWUbKMqaiKvQcI27Mss,5884
@@ -92,9 +94,9 @@ liger_kernel/transformers/trainer/__init__.py,sha256=p7yQfklV8-467qSz_ZMimkbDF7H
 liger_kernel/transformers/trainer/orpo_trainer.py,sha256=tX0h63aOFe3rNqTmk6JpMf75UPo981yzEa6TghnjS0Q,5370
 liger_kernel/triton/__init__.py,sha256=qCiCamzCRv6lpV8IqpAc9YMdNKC7GKurClWceQPnlis,92
 liger_kernel/triton/monkey_patch.py,sha256=Rd0hUHAzDkFfHvnX7-PBaNK5EKnZhtfM_h-fgQH9HPY,1568
-liger_kernel_nightly-0.6.1.dev20250805235815.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel_nightly-0.6.1.dev20250805235815.dist-info/METADATA,sha256=CS11ODzzuxnxvqj1e-waNYiUcvwS5gwkrtGEwYnTkn0,24504
-liger_kernel_nightly-0.6.1.dev20250805235815.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel_nightly-0.6.1.dev20250805235815.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
-liger_kernel_nightly-0.6.1.dev20250805235815.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel_nightly-0.6.1.dev20250805235815.dist-info/RECORD,,
+liger_kernel_nightly-0.6.1.dev20250809233744.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel_nightly-0.6.1.dev20250809233744.dist-info/METADATA,sha256=nClPzQutLCx1b5T1KZGIsCPMTAaYbSw6PVkt9EJjhVw,24504
+liger_kernel_nightly-0.6.1.dev20250809233744.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel_nightly-0.6.1.dev20250809233744.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
+liger_kernel_nightly-0.6.1.dev20250809233744.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel_nightly-0.6.1.dev20250809233744.dist-info/RECORD,,