liger-kernel-nightly 0.6.3.dev20251121010306__py3-none-any.whl → 0.6.4.dev20251121224847__py3-none-any.whl

liger_kernel/chunked_loss/fused_linear_ppo.py

@@ -32,7 +32,7 @@ class LigerFusedLinearPPOBase(torch.autograd.Function):
         epsilon_low=0.2,
         epsilon_high=0.2,
         beta=0.04,
-        loss_type="bnpo",
+        loss_type="dapo",
         max_completion_length=None,
         importance_sampling_level="token",
         temperature=1.0,
@@ -60,7 +60,7 @@ class LigerFusedLinearPPOBase(torch.autograd.Function):
             epsilon_low: Lower bound for clipping the importance sampling ratio
             epsilon_high: Upper bound for clipping the importance sampling ratio
             beta: Weight for the KL penalty
-            loss_type: Type of loss calculation ("grpo", "bnpo", "dr_grpo")
+            loss_type: Type of loss calculation ("grpo", "bnpo", "dr_grpo", "dapo")
             max_completion_length: Maximum completion length required for "dr_grpo"
             temperature: Temperature for the logits
             compiled: Whether to use torch compile
@@ -244,6 +244,21 @@ class LigerFusedLinearPPOBase(torch.autograd.Function):
 
         return loss_acc, tuple(final_metrics)
 
+    @staticmethod
+    def _compute_dapo_normalizer(attention_mask):
+        """Global active tokens averaged per process."""
+        normalizer = attention_mask.to(torch.float32).sum()
+        world_size = 1
+        if torch.distributed.is_available() and torch.distributed.is_initialized():
+            import torch.distributed as dist
+
+            normalizer = normalizer.clone()
+            dist.all_reduce(normalizer, op=dist.ReduceOp.SUM)
+            world_size = dist.get_world_size()
+
+        normalizer = normalizer / world_size
+        return torch.clamp(normalizer, min=1.0)
+
     @staticmethod
     def _compute_chunk_loss(
         input_chunk,
@@ -261,7 +276,7 @@ class LigerFusedLinearPPOBase(torch.autograd.Function):
         epsilon_low=0.2,
         epsilon_high=0.2,
         beta=0.04,
-        loss_type="bnpo",
+        loss_type="dapo",
         max_completion_length=None,
         importance_sampling_level="token",
         temperature=1.0,
@@ -341,10 +356,11 @@ class LigerFusedLinearPPOBase(torch.autograd.Function):
             None,  # grad_epsilon_low
             None,  # grad_epsilon_high
             None,  # grad_beta
+            None,  # grad_loss_type
+            None,  # grad_max_completion_length
+            None,  # grad_importance_sampling_level
             None,  # grad_temperature
             None,  # grad_compiled
             None,  # grad_use_ref_model
             None,  # grad_chunk_size
-            None,  # grad_loss_type
-            None,  # grad_max_completion_length
         )

liger_kernel/chunked_loss/grpo_loss.py

@@ -29,7 +29,7 @@ class LigerFusedLinearGRPOFunction(LigerFusedLinearPPOBase):
         epsilon_low=0.2,
         epsilon_high=0.2,
         beta=0.04,
-        loss_type="bnpo",  # ["grpo", "bnpo", "dr_grpo"]
+        loss_type="dapo",  # ["grpo", "bnpo", "dr_grpo", "dapo"]
         max_completion_length=None,  # Required for dr_grpo
         importance_sampling_level="token",  # ["token", "sequence"] - new parameter for GSPO
         **kwargs,
@@ -94,6 +94,9 @@ class LigerFusedLinearGRPOFunction(LigerFusedLinearPPOBase):
             if max_completion_length is None:
                 raise ValueError("max_completion_length must be provided for loss_type 'dr_grpo'")
             loss = (per_token_loss * attention_mask).sum() / (full_attention_mask.shape[0] * max_completion_length)
+        elif loss_type == "dapo":
+            loss_normalizer = LigerFusedLinearPPOBase._compute_dapo_normalizer(full_attention_mask)
+            loss = (per_token_loss * attention_mask).sum() / loss_normalizer
         else:
             raise ValueError(f"Unknown loss type: {loss_type}")
 
@@ -135,7 +138,7 @@ class LigerFusedLinearGRPOFunction(LigerFusedLinearPPOBase):
         beta=0.04,
         epsilon_low=0.2,
         epsilon_high=0.2,
-        loss_type="bnpo",
+        loss_type="dapo",
         max_completion_length=None,
         importance_sampling_level="token",
         temperature=1.0,
@@ -157,7 +160,7 @@ class LigerFusedLinearGRPOFunction(LigerFusedLinearPPOBase):
             ref_weight (torch.Tensor, optional): Reference model weight tensor. Shape: (vocab_size, hidden_size)
             ref_bias (torch.Tensor, optional): Reference model bias tensor. Shape: (vocab_size,)
             beta (float): Weight for the KL penalty
-            loss_type (str): Type of loss calculation ("grpo", "bnpo", "dr_grpo"). Defaults to "bnpo".
+            loss_type (str): Type of loss calculation ("grpo", "bnpo", "dr_grpo", "dapo"). Defaults to "dapo".
             max_completion_length (int, optional): Maximum completion length, required for "dr_grpo". Defaults to None.
             importance_sampling_level (str): Level of importance sampling ("token" or "sequence"). Defaults to "token".
             temperature (float): Temperature for the logits
@@ -235,7 +238,7 @@ class LigerFusedLinearGRPOLoss(torch.nn.Module):
         chunk_size: int = 1,
         epsilon_low: float = 0.2,
         epsilon_high: float = 0.2,
-        loss_type: str = "bnpo",
+        loss_type: str = "dapo",
         max_completion_length: Optional[int] = None,
         importance_sampling_level: str = "token",
         temperature: float = 1.0,
@@ -248,7 +251,7 @@ class LigerFusedLinearGRPOLoss(torch.nn.Module):
             chunk_size (int): Size of chunks for processing.
             epsilon_low (float): Lower bound for the importance sampling ratio.
             epsilon_high (float): Upper bound for the importance sampling ratio.
-            loss_type (str): Type of loss calculation ("grpo", "bnpo", "dr_grpo"). Defaults to "bnpo".
+            loss_type (str): Type of loss calculation ("grpo", "bnpo", "dr_grpo", "dapo"). Defaults to "dapo".
             max_completion_length (int, optional): Maximum completion length, required for "dr_grpo". Defaults to None.
             importance_sampling_level (str): Level of importance sampling ("token" or "sequence"). Defaults to "token".
             temperature (float): Temperature for the logits.

liger_kernel/ops/grpo_loss.py

@@ -128,7 +128,9 @@ def _grpo_loss_fwd_kernel(
     per_token_loss1 = coef_1 * advantage
     per_token_loss2 = coef_2 * advantage
     per_token_loss = -tl.minimum(per_token_loss1, per_token_loss2)
-    is_clipped = per_token_loss1 < per_token_loss2
+    is_low_clipped = (coef_1 < 1 - EPS_LOW) & (advantage < 0)
+    is_high_clipped = (coef_1 > 1 + EPS_HIGH) & (advantage > 0)
+    is_clipped = is_low_clipped | is_high_clipped
 
     if BETA != 0.0:
         REF_LOGP += off_b * L + off_l

liger_kernel/ops/layer_norm.py

@@ -1,3 +1,4 @@
+import math
 import operator
 
 import torch
@@ -85,68 +86,87 @@ def _layer_norm_forward_kernel(
 @triton.jit
 def _layer_norm_backward_kernel(
     X_ptr,  # pointer to input, shape (n_rows, n_cols)
+    stride_x,  # stride of each row in input
     W_ptr,  # pointer to weights, shape (n_cols,)
     Mean_ptr,  # pointer to mean, shape (n_rows,)
+    stride_mean,  # stride of each row in mean
     RSTD_ptr,  # pointer to rstd, shape (n_rows,)
+    stride_rstd,  # stride of each row in rstd
     DX_ptr,  # pointer to input grad, shape (n_rows, n_cols)
+    stride_dx,  # stride of each row in input grad
     DW_ptr,  # pointer to weights grad, shape (n_cols,)
+    stride_dw,  # stride of each row in weights grad
     DB_ptr,  # pointer to bias grad, shape (n_cols,)
+    stride_db,  # stride of each row in bias grad
     DY_ptr,  # pointer to output grad, shape (n_rows, n_cols)
-    stride_x,  # stride of each row in input
-    stride_dx,  # stride of each row in input grad
     stride_dy,  # stride of each row in output grad
+    n_rows,
     n_cols,
+    rows_per_program: tl.constexpr,
     BLOCK_SIZE: tl.constexpr,
-    dtype: tl.constexpr,
-    atomic_dtype: tl.constexpr,
 ):
     """
     References:
     https://arxiv.org/abs/1607.06450
     https://github.com/karpathy/llm.c/blob/master/doc/layernorm/layernorm.md
     """
-    row_idx = tl.program_id(0).to(tl.int64)
+    row_block_id = tl.program_id(0).to(tl.int64)
+    row_start = row_block_id * rows_per_program
+    row_end = min((row_block_id + 1) * rows_per_program, n_rows)
     cols = tl.arange(0, BLOCK_SIZE)
     mask = cols < n_cols
 
+    dW_row = tl.zeros((BLOCK_SIZE,), dtype=tl.float32)
+    db_row = tl.zeros((BLOCK_SIZE,), dtype=tl.float32)
+
     # Pre-load weights once (same optimization as forward pass)
     w = tl.load(W_ptr + cols, mask=mask, other=0.0)
     w_f32 = w.to(tl.float32)
 
     # Calculate pointers for this specific row
-    row_X_ptr = X_ptr + row_idx * stride_x
-    row_DX_ptr = DX_ptr + row_idx * stride_dx
-    row_DY_ptr = DY_ptr + row_idx * stride_dy
-    row_Mean_ptr = Mean_ptr + row_idx
-    row_RSTD_ptr = RSTD_ptr + row_idx
-
-    # Load data for this row
-    x = tl.load(row_X_ptr + cols, mask=mask, other=0.0)
-    dy = tl.load(row_DY_ptr + cols, mask=mask, other=0.0)
-    mean = tl.load(row_Mean_ptr)
-    rstd = tl.load(row_RSTD_ptr)
-
-    # Convert to fp32 for numerical stability
-    x_f32 = x.to(tl.float32)
-    dy_f32 = dy.to(tl.float32)
-    mean_f32 = mean.to(tl.float32)
-    rstd_f32 = rstd.to(tl.float32)
-
-    # Compute backward pass for this row
-    x_hat = (x_f32 - mean_f32) * rstd_f32
-    wdy = w_f32 * dy_f32
-    c1 = tl.sum(x_hat * wdy, axis=0) / n_cols
-    c2 = tl.sum(wdy, axis=0) / n_cols
-    dx = (wdy - (x_hat * c1 + c2)) * rstd_f32
-
-    # Store input gradient
-    tl.store(row_DX_ptr + cols, dx.to(dtype), mask=mask)
-
-    # Accumulate weight and bias gradients using atomic operations
-    dw = dy_f32 * x_hat
-    db = dy_f32
-    tl.atomic_add(DW_ptr + cols, dw.to(atomic_dtype), mask=mask)
-    tl.atomic_add(DB_ptr + cols, db.to(atomic_dtype), mask=mask)
+    row_X_ptr = X_ptr + row_start * stride_x
+    row_DX_ptr = DX_ptr + row_start * stride_dx
+    row_DY_ptr = DY_ptr + row_start * stride_dy
+    row_Mean_ptr = Mean_ptr + row_start
+    row_RSTD_ptr = RSTD_ptr + row_start
+
+    for _ in range(row_start, row_end):
+        # Load data for this row
+        x = tl.load(row_X_ptr + cols, mask=mask, other=0.0)
+        dy = tl.load(row_DY_ptr + cols, mask=mask, other=0.0)
+        mean = tl.load(row_Mean_ptr)
+        rstd = tl.load(row_RSTD_ptr)
+
+        # Convert to fp32 for numerical stability
+        x_f32 = x.to(tl.float32)
+        dy_f32 = dy.to(tl.float32)
+        mean_f32 = mean.to(tl.float32)
+        rstd_f32 = rstd.to(tl.float32)
+
+        # Compute backward pass for this row
+        x_hat = (x_f32 - mean_f32) * rstd_f32
+        wdy = w_f32 * dy_f32
+        c1 = tl.sum(x_hat * wdy, axis=0) / n_cols
+        c2 = tl.sum(wdy, axis=0) / n_cols
+        dx = (wdy - (x_hat * c1 + c2)) * rstd_f32
+
+        # Store input gradient
+        tl.store(row_DX_ptr + cols, dx, mask=mask)
+
+        # Accumulate weight and bias gradients for this thread block's assigned rows
+        dw = dy_f32 * x_hat
+        db = dy_f32
+        dW_row += dw
+        db_row += db
+
+        row_X_ptr += stride_x
+        row_DX_ptr += stride_dx
+        row_DY_ptr += stride_dy
+        row_Mean_ptr += stride_mean
+        row_RSTD_ptr += stride_rstd
+
+    tl.store(DW_ptr + row_block_id * stride_dw + cols, dW_row, mask=mask)
+    tl.store(DB_ptr + row_block_id * stride_db + cols, db_row, mask=mask)
 
 
 def layer_norm_forward(X, W, B, eps):
@@ -228,31 +248,25 @@ def layer_norm_backward(dY, X, W, B, Mean, RSTD):
     dY = dY.view(-1, dim)
     n_rows, n_cols = dY.shape
 
-    # Allocate gradient tensors
-    DX = torch.empty((n_rows, n_cols), dtype=X.dtype, device=X.device)
-    # Use float32 for weight/bias gradients if bfloat16 (due to atomic_add limitation)
-    grad_dtype = torch.float32 if W.dtype == torch.bfloat16 else W.dtype
-    DW = torch.zeros(n_cols, dtype=grad_dtype, device=W.device)
-    DB = torch.zeros(n_cols, dtype=grad_dtype, device=W.device)
+    sm_count = 1
+    if X.device.type == "cuda":
+        sm_count = torch.cuda.get_device_properties(X.device).multi_processor_count
+    elif X.device.type == "xpu":
+        sm_count = torch.xpu.get_device_properties(X.device).gpu_eu_count
+
+    # fp32 for numerical stability especially.
+    _DW = torch.empty((sm_count, n_cols), dtype=torch.float32, device=W.device)
+    _DB = torch.empty((sm_count, n_cols), dtype=torch.float32, device=W.device)
 
     # Calculate optimal block size and warp configuration
     BLOCK_SIZE, num_warps = calculate_settings(n_cols)
     if n_cols > BLOCK_SIZE:
         raise RuntimeError(f"Feature dimension {n_cols} exceeds maximum supported size of {BLOCK_SIZE}.")
+    rows_per_program = math.ceil(n_rows / sm_count)
+    grid = (sm_count,)
 
-    # Determine dtype for triton operations
-    triton_dtype = (
-        tl.float32
-        if X.dtype == torch.float32
-        else tl.bfloat16
-        if X.dtype == torch.bfloat16
-        else tl.float16
-        if X.dtype == torch.float16
-        else tl.float32  # fallback
-    )
-
-    # Use float32 for atomic operations if bfloat16 is not supported
-    atomic_dtype = tl.float32 if triton_dtype == tl.bfloat16 else triton_dtype
+    # Allocate gradient tensors
+    DX = torch.empty((n_rows, n_cols), dtype=X.dtype, device=X.device)
 
     kernel_args = {"num_warps": num_warps}
     # XPU-specific optimization
@@ -260,28 +274,33 @@ def layer_norm_backward(dY, X, W, B, Mean, RSTD):
         kernel_args.update({"grf_mode": "large", "num_warps": 32, "num_stages": 4})
 
     # Launch kernel with one thread block per row for optimal performance
-    grid = (n_rows,)
     _layer_norm_backward_kernel[grid](
         X,
+        X.stride(0),
         W,
         Mean,
+        Mean.stride(0),
         RSTD,
+        RSTD.stride(0),
         DX,
-        DW,
-        DB,
-        dY,
-        X.stride(0),
         DX.stride(0),
+        _DW,
+        _DW.stride(0),
+        _DB,
+        _DB.stride(0),
+        dY,
         dY.stride(0),
+        n_rows,
         n_cols,
+        rows_per_program=rows_per_program,
         BLOCK_SIZE=BLOCK_SIZE,
-        dtype=triton_dtype,
-        atomic_dtype=atomic_dtype,
         **kernel_args,
     )
 
     DX = DX.view(*shape)
-    return DX, DW.to(W.dtype), DB.to(W.dtype)
+    DW = _DW.sum(dim=0).to(W.dtype)
+    DB = _DB.sum(dim=0).to(B.dtype)
+    return DX, DW, DB
 
 
 class LigerLayerNormFunction(torch.autograd.Function):

liger_kernel/transformers/__init__.py

@@ -42,6 +42,8 @@ if TYPE_CHECKING:
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_glm4v  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_glm4v_moe  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_granite  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_hunyuan_v1_dense  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_hunyuan_v1_moe  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_internvl  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_llama  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_llama4  # noqa: F401
@@ -50,6 +52,7 @@ if TYPE_CHECKING:
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_mixtral  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_mllama  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_olmo2  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_olmo3  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_paligemma  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_phi3  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen2  # noqa: F401
@@ -116,6 +119,7 @@ def __getattr__(name: str):
         "apply_liger_kernel_to_mixtral",
         "apply_liger_kernel_to_mllama",
         "apply_liger_kernel_to_olmo2",
+        "apply_liger_kernel_to_olmo3",
         "apply_liger_kernel_to_paligemma",
         "apply_liger_kernel_to_phi3",
         "apply_liger_kernel_to_qwen2",
@@ -128,6 +132,8 @@ def __getattr__(name: str):
         "apply_liger_kernel_to_qwen3_vl_moe",
         "apply_liger_kernel_to_smollm3",
         "apply_liger_kernel_to_smolvlm",
+        "apply_liger_kernel_to_hunyuan_v1_dense",
+        "apply_liger_kernel_to_hunyuan_v1_moe",
     }
 
     if name in monkey_patch_symbols:
@@ -190,6 +196,7 @@ if _TRANSFORMERS_AVAILABLE:
             "apply_liger_kernel_to_mixtral",
             "apply_liger_kernel_to_mllama",
             "apply_liger_kernel_to_olmo2",
+            "apply_liger_kernel_to_olmo3",
             "apply_liger_kernel_to_paligemma",
             "apply_liger_kernel_to_phi3",
             "apply_liger_kernel_to_qwen2",
@@ -202,5 +209,7 @@ if _TRANSFORMERS_AVAILABLE:
             "apply_liger_kernel_to_qwen3_vl_moe",
             "apply_liger_kernel_to_smollm3",
             "apply_liger_kernel_to_smolvlm",
+            "apply_liger_kernel_to_hunyuan_v1_dense",
+            "apply_liger_kernel_to_hunyuan_v1_moe",
         ]
     )

liger_kernel/transformers/grpo_loss.py

@@ -1,3 +1,6 @@
+import torch
+
+from liger_kernel.chunked_loss.fused_linear_ppo import LigerFusedLinearPPOBase
 from liger_kernel.ops.grpo_loss import GrpoLossFunction
 
 
@@ -13,12 +16,20 @@ def triton_grpo_loss(
     eps_low=0.2,
     eps_high=0.4,
     inplace=True,
+    loss_type="dapo",
+    max_completion_length=None,
+    importance_sampling_level="token",
+    reduce=False,
 ):
     assert logits is not None and completion_ids is not None and advantages is not None, (
         "must provide logits、completion_ids and advantages"
     )
+    if importance_sampling_level != "token":
+        raise ValueError(
+            f"Triton GRPO loss only supports token-level importance sampling. Got {importance_sampling_level}."
+        )
 
-    return GrpoLossFunction.apply(
+    per_token_loss, per_token_kl, is_clipped = GrpoLossFunction.apply(
         logits,
         old_logp,
         ref_logp,
@@ -31,6 +42,50 @@ def triton_grpo_loss(
         eps_high,
         inplace,
     )
+    if not reduce:
+        return per_token_loss, per_token_kl, is_clipped
+
+    loss = _reduce_grpo_loss(
+        per_token_loss,
+        completion_mask,
+        loss_type=loss_type,
+        max_completion_length=max_completion_length,
+    )
+
+    metrics = []
+    if beta != 0.0 and per_token_kl is not None:
+        metrics.append(_masked_mean(per_token_kl, completion_mask))
+    metrics.append(_masked_mean(is_clipped.float(), completion_mask))
+    return loss, metrics
+
+
+def _reduce_grpo_loss(per_token_loss, completion_mask, loss_type, max_completion_length):
+    mask = completion_mask
+    if mask is None:
+        mask = torch.ones_like(per_token_loss, dtype=per_token_loss.dtype, device=per_token_loss.device)
+    mask = mask.to(per_token_loss.dtype)
+
+    if loss_type == "grpo":
+        per_seq = (per_token_loss * mask).sum(-1) / mask.sum(-1).clamp(min=1.0)
+        return per_seq.mean()
+    if loss_type == "bnpo":
+        return (per_token_loss * mask).sum() / mask.sum().clamp(min=1.0)
+    if loss_type == "dr_grpo":
+        if max_completion_length is None:
+            raise ValueError("max_completion_length must be provided when using loss_type='dr_grpo'")
+        batch = per_token_loss.shape[0]
+        return (per_token_loss * mask).sum() / (batch * max_completion_length)
+    if loss_type == "dapo":
+        normalizer = LigerFusedLinearPPOBase._compute_dapo_normalizer(mask)
+        return (per_token_loss * mask).sum() / normalizer
+    raise ValueError(f"Unsupported loss_type '{loss_type}' for Triton GRPO loss.")
+
+
+def _masked_mean(values, mask):
+    if mask is None:
+        mask = torch.ones_like(values, dtype=values.dtype, device=values.device)
+    mask = mask.to(values.dtype)
+    return (values * mask).sum() / mask.sum().clamp(min=1.0)
 
 
 # This is a demo how to use grpo_loss in GRPOTrainer. The Trl version must be 0.16

liger_kernel/transformers/model/hunyuan_v1.py

@@ -0,0 +1,134 @@
+from typing import List
+from typing import Optional
+from typing import Union
+
+import torch
+
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+from liger_kernel.transformers.model.loss_utils import unpack_cross_entropy_result
+from liger_kernel.transformers.model.output_classes import LigerCausalLMOutputWithPast
+
+
+def lce_forward(
+    self,
+    input_ids: Optional[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,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    skip_logits: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    **kwargs,
+) -> LigerCausalLMOutputWithPast:
+    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]`.
+
+        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, HunYuanDenseV1ForCausalLM
+
+    >>> model = HunYuanDenseV1ForCausalLM.from_pretrained("meta-hunyuan_v1_dense/HunYuanDenseV1-2-7b-hf")
+    >>> tokenizer = AutoTokenizer.from_pretrained("meta-hunyuan_v1_dense/HunYuanDenseV1-2-7b-hf")
+
+    >>> 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_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,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    hidden_states = outputs[0]
+    # 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, :]
+
+    shift_labels = kwargs.pop("shift_labels", None)
+    logits = None
+    loss = None
+    token_accuracy = None
+
+    if skip_logits and labels is None and shift_labels is None:
+        raise ValueError("skip_logits is True, but labels and shift_labels are 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 or shift_labels is not None)
+
+    # Compute loss
+    if skip_logits:
+        result = LigerForCausalLMLoss(
+            hidden_states=kept_hidden_states,
+            lm_head_weight=self.lm_head.weight,
+            labels=labels,
+            shift_labels=shift_labels,
+            hidden_size=self.config.hidden_size,
+            **kwargs,
+        )
+        loss, _, token_accuracy = unpack_cross_entropy_result(result)
+
+    else:
+        logits = self.lm_head(kept_hidden_states)
+        if labels is not None or shift_labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                shift_labels=shift_labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        output = ((loss,) + output) if loss is not None else output
+        output = output + (token_accuracy,) if token_accuracy is not None else output
+        return output
+
+    # Return custom output class with accuracy field
+    return LigerCausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        token_accuracy=token_accuracy,
+    )

liger_kernel/transformers/model/olmo3.py

@@ -0,0 +1,142 @@
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+
+from transformers.modeling_outputs import BaseModelOutputWithPast
+from transformers.utils.deprecation import deprecate_kwarg
+
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+from liger_kernel.transformers.model.loss_utils import unpack_cross_entropy_result
+from liger_kernel.transformers.model.output_classes import LigerCausalLMOutputWithPast
+
+
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+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,
+    return_dict: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    skip_logits: Optional[bool] = None,
+    **kwargs,
+) -> Union[Tuple, LigerCausalLMOutputWithPast]:
+    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, Olmo3ForCausalLM
+
+    >>> model = Olmo3ForCausalLM.from_pretrained("allenai/Olmo-3-7B-Instruct")
+    >>> tokenizer = AutoTokenizer.from_pretrained("allenai/Olmo-3-7B-Instruct")
+
+    >>> 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 sure if you’re conscious of this, but I’m'
+    ```
+    """
+    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: 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]
+    # 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, :]
+
+    shift_labels = kwargs.pop("shift_labels", None)
+    logits = None
+    loss = None
+    token_accuracy = None
+
+    if skip_logits and labels is None and shift_labels is None:
+        raise ValueError("skip_logits is True, but labels and shift_labels are 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 or shift_labels is not None)
+
+    # Compute loss
+    if skip_logits:
+        result = LigerForCausalLMLoss(
+            hidden_states=kept_hidden_states,
+            lm_head_weight=self.lm_head.weight,
+            labels=labels,
+            shift_labels=shift_labels,
+            hidden_size=self.config.hidden_size,
+            **kwargs,
+        )
+        loss, _, token_accuracy = unpack_cross_entropy_result(result)
+
+    else:
+        logits = self.lm_head(kept_hidden_states)
+        if labels is not None or shift_labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                shift_labels=shift_labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        output = ((loss,) + output) if loss is not None else output
+        output = output + (token_accuracy,) if token_accuracy is not None else output
+        return output
+
+    # Return custom output class with token_accuracy field
+    return LigerCausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        token_accuracy=token_accuracy,
+    )

liger_kernel/transformers/monkey_patch.py

@@ -1928,6 +1928,74 @@ def apply_liger_kernel_to_olmo2(
                 _patch_rms_norm_module(decoder_layer.post_feedforward_layernorm, in_place=False)
 
 
+def apply_liger_kernel_to_olmo3(
+    rope: bool = True,
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = True,
+    rms_norm: bool = True,
+    swiglu: bool = True,
+    model: PreTrainedModel = None,
+) -> None:
+    """
+    Apply Liger kernels to replace original implementation in HuggingFace Olmo3 models.
+
+    Args:
+        rope (bool): Whether to apply Liger's rotary position embedding. Default is True.
+        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 linear 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.
+        swiglu (bool): Whether to apply Liger's SwiGLU to Olmo3MLP. Default is True.
+        model (PreTrainedModel): The model instance to apply Liger kernels to, if the model has already been
+        loaded. Default is None.
+    """
+    assert not (cross_entropy and fused_linear_cross_entropy), (
+        "cross_entropy and fused_linear_cross_entropy cannot both be True."
+    )
+
+    from transformers.models.olmo3 import modeling_olmo3
+    from transformers.models.olmo3.modeling_olmo3 import Olmo3Model
+
+    from liger_kernel.transformers.model.olmo3 import lce_forward as olmo3_lce_forward
+    from liger_kernel.transformers.rms_norm import LigerRMSNormForOlmo2
+
+    # Olmo3 arch is very similar to Olmo2, so we can reuse all these components in the same way.
+    if rope:
+        modeling_olmo3.apply_rotary_pos_emb = liger_rotary_pos_emb
+    if rms_norm:
+        modeling_olmo3.Olmo3RMSNorm = LigerRMSNormForOlmo2  # same as olmo2
+    if swiglu:
+        modeling_olmo3.Olmo3MLP = LigerSwiGLUMLP
+    if cross_entropy:
+        from transformers.loss.loss_utils import nn
+
+        nn.functional.cross_entropy = liger_cross_entropy
+    if fused_linear_cross_entropy:
+        if model is not None:
+            model.forward = MethodType(olmo3_lce_forward, model)
+        else:
+            modeling_olmo3.Olmo3ForCausalLM.forward = olmo3_lce_forward
+
+    if model is not None:
+        # The model instance already exists, so we need to additionally patch the
+        # instance variables that reference already-instantiated modules
+
+        # get the base model from the model instance
+        base_model: Olmo3Model = getattr(model, model.base_model_prefix, model)
+
+        if rms_norm:
+            _patch_rms_norm_module(base_model.norm)
+
+        for decoder_layer in base_model.layers:
+            if swiglu:
+                _patch_swiglu_module(decoder_layer.mlp, LigerSwiGLUMLP)
+            if rms_norm:
+                _patch_rms_norm_module(decoder_layer.post_attention_layernorm, in_place=False)
+                _patch_rms_norm_module(decoder_layer.post_feedforward_layernorm, in_place=False)
+
+
 def apply_liger_kernel_to_glm4(
     rope: bool = False,
     cross_entropy: bool = False,
@@ -2558,6 +2626,123 @@ def apply_liger_kernel_to_qwen3_next(
                             _patch_swiglu_module(expert, LigerQwen3MoeSwiGLUMLP)
 
 
+def apply_liger_kernel_to_hunyuan_v1_dense(
+    rope: bool = True,
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = True,
+    rms_norm: bool = True,
+    swiglu: bool = True,
+    model: PreTrainedModel = None,
+) -> None:
+    """
+    Apply Liger kernels to replace original implementation in HuggingFace Hunyuan v1 dense models.
+    """
+    assert not (cross_entropy and fused_linear_cross_entropy), (
+        "cross_entropy and fused_linear_cross_entropy cannot both be True."
+    )
+
+    from transformers.models.hunyuan_v1_dense import modeling_hunyuan_v1_dense
+    from transformers.models.hunyuan_v1_dense.modeling_hunyuan_v1_dense import HunYuanDenseV1Model
+
+    from liger_kernel.transformers.model.hunyuan_v1 import lce_forward as hunyuan_v1_lce_forward
+    from liger_kernel.transformers.swiglu import LigerHunyuanV1SwiGLUMLP
+
+    if rope:
+        modeling_hunyuan_v1_dense.apply_rotary_pos_emb = liger_rotary_pos_emb
+
+    if rms_norm:
+        modeling_hunyuan_v1_dense.HunYuanDenseV1RMSNorm = LigerRMSNorm
+
+    if cross_entropy:
+        from transformers.loss.loss_utils import nn
+
+        nn.functional.cross_entropy = liger_cross_entropy
+
+    if fused_linear_cross_entropy:
+        if model is not None:
+            model.forward = MethodType(hunyuan_v1_lce_forward, model)
+        else:
+            modeling_hunyuan_v1_dense.HunYuanDenseV1ForCausalLM.forward = hunyuan_v1_lce_forward
+
+    if swiglu:
+        modeling_hunyuan_v1_dense.HunYuanDenseV1MLP = LigerHunyuanV1SwiGLUMLP
+
+    if model is not None:
+        # The model instance already exists, so we need to additionally patch the
+        # instance variables that reference already-instantiated modules
+
+        # get the base model from the model instance
+        base_model: HunYuanDenseV1Model = getattr(model, model.base_model_prefix, model)
+
+        if rms_norm:
+            _patch_rms_norm_module(base_model.norm)
+        for decoder_layer in base_model.layers:
+            if swiglu:
+                _patch_swiglu_module(decoder_layer.mlp, LigerHunyuanV1SwiGLUMLP)
+            if rms_norm:
+                _patch_rms_norm_module(decoder_layer.input_layernorm)
+                _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
+
+
+def apply_liger_kernel_to_hunyuan_v1_moe(
+    rope: bool = True,
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = True,
+    rms_norm: bool = True,
+    swiglu: bool = True,
+    model: PreTrainedModel = None,
+) -> None:
+    """
+    Apply Liger kernels to replace original implementation in HuggingFace Qwen3 models.
+    """
+    assert not (cross_entropy and fused_linear_cross_entropy), (
+        "cross_entropy and fused_linear_cross_entropy cannot both be True."
+    )
+
+    from transformers.models.hunyuan_v1_moe import modeling_hunyuan_v1_moe
+    from transformers.models.hunyuan_v1_moe.modeling_hunyuan_v1_moe import HunYuanMoEV1Model
+
+    from liger_kernel.transformers.model.hunyuan_v1 import lce_forward as hunyuan_v1_moe_lce_forward
+    from liger_kernel.transformers.swiglu import LigerHunyuanV1SwiGLUMLP
+
+    if rope:
+        modeling_hunyuan_v1_moe.apply_rotary_pos_emb = liger_rotary_pos_emb
+
+    if rms_norm:
+        modeling_hunyuan_v1_moe.HunYuanMoEV1RMSNorm = LigerRMSNorm
+
+    if cross_entropy:
+        from transformers.loss.loss_utils import nn
+
+        nn.functional.cross_entropy = liger_cross_entropy
+
+    if fused_linear_cross_entropy:
+        if model is not None:
+            model.forward = MethodType(hunyuan_v1_moe_lce_forward, model)
+        else:
+            modeling_hunyuan_v1_moe.HunYuanMoEV1ForCausalLM.forward = hunyuan_v1_moe_lce_forward
+
+    if swiglu:
+        modeling_hunyuan_v1_moe.HunYuanMoEV1MLP = LigerHunyuanV1SwiGLUMLP
+
+    if model is not None:
+        # The model instance already exists, so we need to additionally patch the
+        # instance variables that reference already-instantiated modules
+
+        # get the base model from the model instance
+        base_model: HunYuanMoEV1Model = getattr(model, model.base_model_prefix, model)
+
+        if rms_norm:
+            _patch_rms_norm_module(base_model.norm)
+        for decoder_layer in base_model.layers:
+            if swiglu:
+                for mlp_expert in decoder_layer.mlp.experts:
+                    _patch_swiglu_module(mlp_expert, LigerHunyuanV1SwiGLUMLP)
+            if rms_norm:
+                _patch_rms_norm_module(decoder_layer.input_layernorm)
+                _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
+
+
 # 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,
@@ -2578,6 +2763,7 @@ MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "mistral": apply_liger_kernel_to_mistral,
     "mixtral": apply_liger_kernel_to_mixtral,
     "olmo2": apply_liger_kernel_to_olmo2,
+    "olmo3": apply_liger_kernel_to_olmo3,
     "qwen2": apply_liger_kernel_to_qwen2,
     "qwen3": apply_liger_kernel_to_qwen3,
     "qwen3_moe": apply_liger_kernel_to_qwen3_moe,
@@ -2595,6 +2781,8 @@ MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "paligemma": apply_liger_kernel_to_paligemma,
     "falcon_h1": apply_liger_kernel_to_falcon_h1,
     "smolvlm": apply_liger_kernel_to_smolvlm,
+    "hunyuan_v1_dense": apply_liger_kernel_to_hunyuan_v1_dense,
+    "hunyuan_v1_moe": apply_liger_kernel_to_hunyuan_v1_moe,
 }
 
 

liger_kernel/transformers/swiglu.py

@@ -77,3 +77,20 @@ class LigerQwen3MoeSwiGLUMLP(nn.Module):
 
     def forward(self, x):
         return self.down_proj(LigerSiLUMulFunction.apply(self.gate_proj(x), self.up_proj(x)))
+
+
+class LigerHunyuanV1SwiGLUMLP(nn.Module):
+    def __init__(self, config, layer_idx=None, is_shared_mlp=False):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = 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.layer_idx = layer_idx
+        if config.hidden_act not in ["silu", "swish"]:
+            raise ValueError(f"Activation function {config.hidden_act} not supported.")
+
+    def forward(self, x):
+        return self.down_proj(LigerSiLUMulFunction.apply(self.gate_proj(x), self.up_proj(x)))

{liger_kernel_nightly-0.6.3.dev20251121010306.dist-info → liger_kernel_nightly-0.6.4.dev20251121224847.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.3.dev20251121010306
+Version: 0.6.4.dev20251121224847
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -310,8 +310,11 @@ loss.backward()
 | Phi3 & Phi3.5       | `liger_kernel.transformers.apply_liger_kernel_to_phi3`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
 | Granite 3.0 & 3.1   | `liger_kernel.transformers.apply_liger_kernel_to_granite`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss |
 | OLMo2   | `liger_kernel.transformers.apply_liger_kernel_to_olmo2`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
+| Olmo3   | `liger_kernel.transformers.apply_liger_kernel_to_olmo3`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
 | GLM-4   | `liger_kernel.transformers.apply_liger_kernel_to_glm4`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
 | InternVL3   | `liger_kernel.transformers.apply_liger_kernel_to_internvl`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
+| HunyuanV1   | `liger_kernel.transformers.apply_liger_kernel_to_hunyuan_v1_dense`    |  RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy       |
+| HunyuanV1 MoE | `liger_kernel.transformers.apply_liger_kernel_to_hunyuan_v1_moe` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy       |
 
 
 ## Low-level APIs

{liger_kernel_nightly-0.6.3.dev20251121010306.dist-info → liger_kernel_nightly-0.6.4.dev20251121224847.dist-info}/RECORD

@@ -8,10 +8,10 @@ liger_kernel/chunked_loss/cpo_loss.py,sha256=Gzz1eU4kgcbdubFVRy55e8A1Cr-r45UgNic
 liger_kernel/chunked_loss/dpo_loss.py,sha256=I83khNs3QQjuhr8U3NIOAACkbse6DNiBV-TulPZ0lXw,9006
 liger_kernel/chunked_loss/functional.py,sha256=-XPDbLml9dHmvoSU2VNTUrBDFehuzvuAGPikVetBMtI,1132
 liger_kernel/chunked_loss/fused_linear_distillation.py,sha256=yRtolfFGfKB-SxGQQyF68GYXd11Zlvh1InLdGeWNFIE,12652
-liger_kernel/chunked_loss/fused_linear_ppo.py,sha256=ZjpNP5VC-tXXIKb4AckkQ3iWWQeej-JoG4StJq3N0wg,13650
+liger_kernel/chunked_loss/fused_linear_ppo.py,sha256=baU19PwqO1FTVxwlB-eyJv6gOLtL7baXGzSncYQ8Ktc,14296
 liger_kernel/chunked_loss/fused_linear_preference.py,sha256=FIH85uUXAOgYx5Ax8MjFhJHVu-2pKtY7wSegd0zSyyY,18336
 liger_kernel/chunked_loss/fused_linear_unpaired_preference.py,sha256=RiuK3UtRwH9T6jZ36sA8Urj-TVuOLOO2syLg_JOQapY,13437
-liger_kernel/chunked_loss/grpo_loss.py,sha256=SkZuKoW8K94UbWR-OtfopsQkuQ8tFOr_90AGR6_Mhes,12844
+liger_kernel/chunked_loss/grpo_loss.py,sha256=bmuZaNgqNbJ5pJGFDXWE-B4BGYF7xWVSN15UyCfuq_s,13079
 liger_kernel/chunked_loss/jsd_loss.py,sha256=G0RghPYYelyZ6DOEiwS8we9TT5MY2iHpiFqzZ2Xy87g,8038
 liger_kernel/chunked_loss/kto_loss.py,sha256=llVCe6DkcpCo57seGWoMikaQVFApx764jsmSbQyqwQY,7529
 liger_kernel/chunked_loss/orpo_loss.py,sha256=nu9UYG16dcMw93lvHi4_hYs3Q0FK1KnlmMRj7OpYU8s,4872
@@ -25,10 +25,10 @@ liger_kernel/ops/fused_linear_jsd.py,sha256=CSoprxb-YcJy-YUKiTcYkxN8sb9h2kdk_iHu
 liger_kernel/ops/fused_neighborhood_attention.py,sha256=vPi5xbnh6wxyZehaqo6Tuilqo2fN5SGDiONjnNmIKqs,35556
 liger_kernel/ops/geglu.py,sha256=r0WSq9E93zzynL44Wh8femzOWK07_SseBM_pJUyxT3s,4144
 liger_kernel/ops/group_norm.py,sha256=qD4D4lSjSgVtO52EBNLC2iTseALRgPgqXE50U2woggk,10837
-liger_kernel/ops/grpo_loss.py,sha256=anRnv7k1-AV3pCC6_TqP0GMg78YYUfRAJrbpx6PVhl0,9448
+liger_kernel/ops/grpo_loss.py,sha256=2SyOujtF9I3xiNo4wFf4s6MeiDotE_qeYfRWgj_bOBE,9573
 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=WmiORsIyufOhazmYZTPjeSc5Z-xTAYwXAKqUcCv_dlY,9807
+liger_kernel/ops/layer_norm.py,sha256=OMaex1MDsM9kaFs0-q5Pnx3DrMVjongQoZ5-iFIOy00,10523
 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/poly_norm.py,sha256=MLgI8Ea93fugKibHCUauQ2ASYVXCvpPZe5v3kQZU6po,11152
@@ -43,7 +43,7 @@ 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=XX1ySRgZXeQe0or-6GNclAsNQG_VkABQlkwqpB1Wn8A,10090
+liger_kernel/transformers/__init__.py,sha256=CgwhrY5cdx6OcRgR2ZZJbOIkLswQWPTr-BAaoxDNNOY,10687
 liger_kernel/transformers/auto_model.py,sha256=0qCTRZt280Bj_LcFdzo9hlaR-BWNazawXOGgoCZjgEg,1545
 liger_kernel/transformers/cross_entropy.py,sha256=DMtHkKrVJDSsels7KgGQJqrXkEAd6Zopcdr-5oRmQgE,2010
 liger_kernel/transformers/dyt.py,sha256=i-4GPaMrl-jab9TVI5qN0-H9qycn_mCbV82ozU4nbmU,723
@@ -55,12 +55,12 @@ liger_kernel/transformers/fused_linear_jsd.py,sha256=bZ4otCvWBuOnA5XdQL-FzZVItJl
 liger_kernel/transformers/fused_neighborhood_attention.py,sha256=TxYDUAt9B6WSP14aJP66C_2Mbds2sSIPGnamhUSTrC8,7957
 liger_kernel/transformers/geglu.py,sha256=mrgqzIUVd6lN7fkDKLkw5YaESDxDtFgbot430WwPVOQ,1107
 liger_kernel/transformers/group_norm.py,sha256=6qMAWOprr4SzP0YhNVNGQIBpM5aUHplUD2VuGJrMBz0,2173
-liger_kernel/transformers/grpo_loss.py,sha256=uAkUNKSnUGEOqa82L9w2e6AI1kcmG8K45-QxyaT8zhM,3897
+liger_kernel/transformers/grpo_loss.py,sha256=QS6Ycct1E2yMfqoHPBa2sUAu5cmweNPK_-Q_KJE8hb4,6098
 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/llama4_rope.py,sha256=kS6PSHEwf3dS7hD7C7p8S0geugx2EMCiP0h0F7LsUoY,3639
-liger_kernel/transformers/monkey_patch.py,sha256=ZGnLygHuCiKGd6hT-C0pt1aY85f6GNFdV98oCDpxHHo,124742
+liger_kernel/transformers/monkey_patch.py,sha256=4LV6LSz_AAop6HWk1spZm1QigPN9nUDPJu9tK21-jIo,132446
 liger_kernel/transformers/multi_token_attention.py,sha256=K3NIY9_5TPgZ4_Rahn0xnkMXxD_fmlJHK4CWGYvGQp0,1752
 liger_kernel/transformers/poly_norm.py,sha256=g5tC75i3qy1_N26ZUP-jfpct7ivQAEdJfIfx8IXzeyE,1377
 liger_kernel/transformers/qwen2vl_mrope.py,sha256=5EwSqrMdsL9MYspeBMXBsNJKvH0MOmRrtJXAJlnnlOI,1047
@@ -68,7 +68,7 @@ liger_kernel/transformers/rms_norm.py,sha256=HwddVqrqS58jE-M2_4NkFGARtCDBhGnkKyj
 liger_kernel/transformers/rope.py,sha256=VMlDZI6zss9mLaLcN5XCE_ktmYRwAi_Eh4TIgO6NrIQ,2361
 liger_kernel/transformers/softmax.py,sha256=yadlAgE4V2JByMwrDDa2s5SUBp8Jgd57xwnVvAWoBaI,264
 liger_kernel/transformers/sparsemax.py,sha256=0lQA0UEOs4mu8CMruZ3VLhImxQVXJWhPsAKUsYA7vj8,403
-liger_kernel/transformers/swiglu.py,sha256=LZ8YeLIdv2k46JleZMjzubGk98smt6t780kSgcVLsQk,3454
+liger_kernel/transformers/swiglu.py,sha256=dRR69wDWSWfdjtnsTECyxQqWVo5QkdXdXm9SpSQ4Jvw,4291
 liger_kernel/transformers/tiled_mlp.py,sha256=J51-kpzwikDMMhT5bX-RZCKMaXBK6zZc1bhgRYTK5F0,4651
 liger_kernel/transformers/trainer_integration.py,sha256=W3ON51O5GkyzNJsItz0y5rKx-uy2f2cFfveZpqbUdhw,123
 liger_kernel/transformers/tvd.py,sha256=XrRfyJIqN6HFxXk8MYyFVZM1OLz3mtSbRZvWfZ_JerQ,450
@@ -82,6 +82,7 @@ liger_kernel/transformers/model/gemma3.py,sha256=mEV3Kuy-dqfTk_b899Vb-InuD4_DvwH
 liger_kernel/transformers/model/glm4.py,sha256=bSp22iPIjsli4-c_usUOsyh1Bs2gIK8X6ynS0azseUs,5900
 liger_kernel/transformers/model/glm4v.py,sha256=dd-BQpccDCp1SbIxcJ5rG8xcwYQK3KOv1Tgm9TGnZc4,6594
 liger_kernel/transformers/model/glm4v_moe.py,sha256=zKhMdOOrRhlrvCSFaeVYfddL1ubpY8edEO91TN81n98,7135
+liger_kernel/transformers/model/hunyuan_v1.py,sha256=MJvP9xkUFePIV0HLETJM4YPbVCEPkAE1ZI5Jxyiebh0,5731
 liger_kernel/transformers/model/internvl.py,sha256=OOutracs9qrPHSU7FVYar08yinvGrHQVPvo39JEws6w,6473
 liger_kernel/transformers/model/llama.py,sha256=kqZeONzwTBzudoChlKMzq1w23BtYGbxWZC1l1V__JTw,13410
 liger_kernel/transformers/model/llama4.py,sha256=PfkynGVI0xxMs3EtyYpCgaALI6stu25OIrTIymE-pvg,4853
@@ -91,6 +92,7 @@ liger_kernel/transformers/model/mistral.py,sha256=OcwOzVDMwwDbVccVPv-AaocznzWwzL
 liger_kernel/transformers/model/mixtral.py,sha256=YcBDoTEJDgLFJ_RTo180DYGxR8D5Ad9-idumif7kCPE,12130
 liger_kernel/transformers/model/mllama.py,sha256=vAHwCm63sn4kpAY0rDGf_N0HR7KRTBVpBYDVTPOaZTg,12079
 liger_kernel/transformers/model/olmo2.py,sha256=-h2bUOeuPfY1MdShdRvq5_wFDHKP4PEimgIl0fL-BT4,5902
+liger_kernel/transformers/model/olmo3.py,sha256=k2zYOlS8U_b5MwjdToB3tDRQ0bH_mWapVQqJcH8-qAo,6007
 liger_kernel/transformers/model/output_classes.py,sha256=0BGXVR4dYQpSHLkSqpRoXuHMryrceGSlTYRu6pvd8ZY,4542
 liger_kernel/transformers/model/paligemma.py,sha256=r0smHLADkEwfLS6d6ArWoSWEeLt2d_8pmgOO5F04b1o,20793
 liger_kernel/transformers/model/phi3.py,sha256=PT7Kw6yySg-7TsssWfi82eVMN3SWujCqzCqHigAdfeQ,4574
@@ -108,9 +110,9 @@ 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-0.6.3.dev20251121010306.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel_nightly-0.6.3.dev20251121010306.dist-info/METADATA,sha256=HgCaZORIkj1lSLvj1vsjOJ0r9ouWZ-lqPCQ3JrJJMFU,24777
-liger_kernel_nightly-0.6.3.dev20251121010306.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel_nightly-0.6.3.dev20251121010306.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
-liger_kernel_nightly-0.6.3.dev20251121010306.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel_nightly-0.6.3.dev20251121010306.dist-info/RECORD,,
+liger_kernel_nightly-0.6.4.dev20251121224847.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel_nightly-0.6.4.dev20251121224847.dist-info/METADATA,sha256=arayRD-HK6tYJLHubc_dtT2TnAojZ_d13YaPVi-txuQ,25238
+liger_kernel_nightly-0.6.4.dev20251121224847.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel_nightly-0.6.4.dev20251121224847.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
+liger_kernel_nightly-0.6.4.dev20251121224847.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel_nightly-0.6.4.dev20251121224847.dist-info/RECORD,,