liger-kernel-nightly 0.6.2.dev20251011154226__py3-none-any.whl → 0.6.2.dev20251013144132__py3-none-any.whl

liger_kernel/ops/cross_entropy.py

@@ -45,6 +45,7 @@ def liger_cross_entropy_kernel(
     BLOCK_SIZE: tl.constexpr,
     HAS_WEIGHT: tl.constexpr,
     HAS_SOFTCAPPING: tl.constexpr,
+    HAS_GRADIENTS: tl.constexpr,
 ):
     """
     This kernel computes both cross entropy loss and the gradient of the input.
@@ -72,6 +73,7 @@ def liger_cross_entropy_kernel(
     BLOCK_SIZE (int): The block size for Triton operations.
     HAS_WEIGHT (bool): The boolean value to determine whether assigning weight to each of the classes.
     HAS_SOFTCAPPING (bool): The boolean value to determine whether applying soft-capping or not.
+    HAS_GRADIENTS (bool): The boolean value to determine whether calculating gradients in forward pass.
     """
 
     # https://github.com/triton-lang/triton/issues/1058
@@ -155,58 +157,58 @@ def liger_cross_entropy_kernel(
     # For 'sum' reduction, no normalization is applied:
     # dx_y = softmax(x_y) - 1
     # dx_i = softmax(x_i), for i ≠ y
-
-    for i in range(0, n_cols, BLOCK_SIZE):
-        X_offsets = i + tl.arange(0, BLOCK_SIZE)
-        X_block = tl.load(
-            X_ptr + X_offsets,
-            mask=X_offsets < n_cols,
-            other=float("-inf"),
-            # Ensure float32 precision for softmax calculation
-        ).cast(tl.float32)
-        if HAS_SOFTCAPPING:
-            intermediate = tanh(X_block / softcap)
-            X_block = softcap * intermediate
-
-        if not HAS_WEIGHT:
-            # softmax(x_i)
-            X_block = tl.exp(X_block - m) / d
-            # derivative of z-loss: 2 * lse_square_scale * lse * softmax(x_i)
-            X_block += 2 * lse_square_scale * lse * X_block
-            # smoothing term
-            X_block += -eps
-            # special handle dx_y
-            X_block = tl.where(X_offsets != y, X_block, X_block - (1 - label_smoothing))
-            # reduction scale
-            if reduction == "mean":
-                X_block = X_block / n_non_ignore
-        else:
-            weight_block = tl.load(weight_ptr + X_offsets, mask=X_offsets < n_cols)
-            softmax_X = tl.exp(X_block - m) / d
-            # derivative of original_loss
-            dloss_ori = (1 - label_smoothing) * softmax_X
-            # specially handle dx_y
-            dloss_ori = tl.where(X_offsets != y, dloss_ori, dloss_ori - (1 - label_smoothing))
-            dloss_ori = dloss_ori * weight_y
-            # derivative of smooth_loss
-            dloss_smooth = eps * (-weight_block + softmax_X * weight_sum)
-            # derivative of z-loss
-            dz_loss = 2 * lse_square_scale * lse * softmax_X
-            # reduction scale
-            if reduction == "mean":
-                dloss_ori = dloss_ori / sum_non_ignore_weight
-                dloss_smooth = dloss_smooth / sum_non_ignore_weight
-                # TODO: Implement weighted z_loss. Currently, z_loss is not scaled by weight.
-                dz_loss = dz_loss / n_non_ignore
-            # derivative of total_loss
-            X_block = dloss_ori + dloss_smooth + dz_loss
-
-        # chain rule softcapping
-        # d(softcap * tanh(x / softcap)) = (1 - tanh^2(x / softcap))
-        if HAS_SOFTCAPPING:
-            X_block = X_block * (1 - intermediate * intermediate)
-
-        tl.store(X_ptr + X_offsets, X_block, mask=X_offsets < n_cols)
+    if HAS_GRADIENTS:
+        for i in range(0, n_cols, BLOCK_SIZE):
+            X_offsets = i + tl.arange(0, BLOCK_SIZE)
+            X_block = tl.load(
+                X_ptr + X_offsets,
+                mask=X_offsets < n_cols,
+                other=float("-inf"),
+                # Ensure float32 precision for softmax calculation
+            ).cast(tl.float32)
+            if HAS_SOFTCAPPING:
+                intermediate = tanh(X_block / softcap)
+                X_block = softcap * intermediate
+
+            if not HAS_WEIGHT:
+                # softmax(x_i)
+                X_block = tl.exp(X_block - m) / d
+                # derivative of z-loss: 2 * lse_square_scale * lse * softmax(x_i)
+                X_block += 2 * lse_square_scale * lse * X_block
+                # smoothing term
+                X_block += -eps
+                # special handle dx_y
+                X_block = tl.where(X_offsets != y, X_block, X_block - (1 - label_smoothing))
+                # reduction scale
+                if reduction == "mean":
+                    X_block = X_block / n_non_ignore
+            else:
+                weight_block = tl.load(weight_ptr + X_offsets, mask=X_offsets < n_cols)
+                softmax_X = tl.exp(X_block - m) / d
+                # derivative of original_loss
+                dloss_ori = (1 - label_smoothing) * softmax_X
+                # specially handle dx_y
+                dloss_ori = tl.where(X_offsets != y, dloss_ori, dloss_ori - (1 - label_smoothing))
+                dloss_ori = dloss_ori * weight_y
+                # derivative of smooth_loss
+                dloss_smooth = eps * (-weight_block + softmax_X * weight_sum)
+                # derivative of z-loss
+                dz_loss = 2 * lse_square_scale * lse * softmax_X
+                # reduction scale
+                if reduction == "mean":
+                    dloss_ori = dloss_ori / sum_non_ignore_weight
+                    dloss_smooth = dloss_smooth / sum_non_ignore_weight
+                    # TODO: Implement weighted z_loss. Currently, z_loss is not scaled by weight.
+                    dz_loss = dz_loss / n_non_ignore
+                # derivative of total_loss
+                X_block = dloss_ori + dloss_smooth + dz_loss
+
+            # chain rule softcapping
+            # d(softcap * tanh(x / softcap)) = (1 - tanh^2(x / softcap))
+            if HAS_SOFTCAPPING:
+                X_block = X_block * (1 - intermediate * intermediate)
+
+            tl.store(X_ptr + X_offsets, X_block, mask=X_offsets < n_cols)
 
     # We need tl.debug_barrier() to ensure the new result of X_ptr is written as mentioned in
     # https://github.com/triton-lang/triton/blob/ba42a5c68fd0505f8c42f4202d53be0f8d9a5fe0/python/triton/ops/cross_entropy.py#L34
@@ -332,6 +334,7 @@ def cross_entropy_forward(
         BLOCK_SIZE=BLOCK_SIZE,
         HAS_WEIGHT=True if weight is not None else False,
         HAS_SOFTCAPPING=True if softcap is not None else False,
+        HAS_GRADIENTS=_input.requires_grad,
         # TODO: 32 seems to give the best performance
         # Performance is quite sensitive to num_warps
         num_warps=32 if not is_hip() else 16,

liger_kernel/ops/fused_linear_cross_entropy.py

@@ -150,6 +150,7 @@ def fused_linear_cross_entropy_forward(
             RETURN_Z_LOSS=return_z_loss,
             HAS_WEIGHT=True if ce_weight is not None else False,
             HAS_SOFTCAPPING=True if softcap is not None else False,
+            HAS_GRADIENTS=_input.requires_grad,
             BLOCK_SIZE=BLOCK_SIZE,
             num_warps=32 if not is_hip() else 16,
         )
@@ -173,10 +174,10 @@ def fused_linear_cross_entropy_forward(
 
         grad_input[start_idx:end_idx] = grad_logits_chunk @ weight
 
-        if grad_weight is not None:
+        if grad_weight is not None and _input.requires_grad:
             grad_weight += torch.mm(grad_logits_chunk.t(), _input_chunk).float()
 
-        if bias is not None:
+        if bias is not None and _input.requires_grad:
             torch.add(
                 input=grad_bias,
                 other=grad_logits_chunk.sum(dim=0),

liger_kernel/transformers/monkey_patch.py

@@ -469,7 +469,7 @@ def apply_liger_kernel_to_llama4(
             `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 MLP. Default is False.
+        swiglu (bool): Whether to apply Liger's SwiGLU MLP. Default is True.
         model (PreTrainedModel): The model instance to apply Liger kernels to, if the model has already been
         loaded. Default is None.
     """
@@ -522,7 +522,10 @@ def apply_liger_kernel_to_llama4(
                 _patch_rms_norm_module(text_model.norm)
             for decoder_layer in text_model.layers:
                 if swiglu:
-                    _patch_swiglu_module(decoder_layer.feed_forward, LigerSwiGLUMLP)
+                    if decoder_layer.is_moe_layer:
+                        _patch_swiglu_module(decoder_layer.feed_forward.shared_expert, LigerSwiGLUMLP)
+                    else:
+                        _patch_swiglu_module(decoder_layer.feed_forward, LigerSwiGLUMLP)
                 if rms_norm:
                     _patch_rms_norm_module(decoder_layer.input_layernorm)
                     _patch_rms_norm_module(decoder_layer.post_attention_layernorm)

{liger_kernel_nightly-0.6.2.dev20251011154226.dist-info → liger_kernel_nightly-0.6.2.dev20251013144132.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.2.dev20251011154226
+Version: 0.6.2.dev20251013144132
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.6.2.dev20251011154226.dist-info → liger_kernel_nightly-0.6.2.dev20251013144132.dist-info}/RECORD

@@ -17,10 +17,10 @@ liger_kernel/chunked_loss/kto_loss.py,sha256=llVCe6DkcpCo57seGWoMikaQVFApx764jsm
 liger_kernel/chunked_loss/orpo_loss.py,sha256=nu9UYG16dcMw93lvHi4_hYs3Q0FK1KnlmMRj7OpYU8s,4872
 liger_kernel/chunked_loss/simpo_loss.py,sha256=fy2w8KbhMrBv7b1jdIeH3bBFxY52bPQPZb3KwBvmurM,5385
 liger_kernel/ops/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-liger_kernel/ops/cross_entropy.py,sha256=e8THGnhOcy_0SbOLABx67HEM7-B8a8pG7nDKbCRpQKM,19123
+liger_kernel/ops/cross_entropy.py,sha256=OVkani9JEmCJ8IHN3UgJKzGW7zxJWDwy1EaWVcbShgQ,19517
 liger_kernel/ops/dyt.py,sha256=gCLz4S8aul8SY9nvIGaoK67aGb7U9MJRQdo3ONqmQYs,5417
 liger_kernel/ops/fused_add_rms_norm.py,sha256=UBqmlqFCmhSAIpkNKd8rrfXatX7Z4J9bp2dX9A0lrJQ,14017
-liger_kernel/ops/fused_linear_cross_entropy.py,sha256=6rB3pdwU97Ivl2IHndPJjzhP28E9Fd0pUQcPHLiuCjc,14290
+liger_kernel/ops/fused_linear_cross_entropy.py,sha256=PqIPHU8EjkHRJF6cNZViDucFVOgqo7eanJxB53Npke8,14388
 liger_kernel/ops/fused_linear_jsd.py,sha256=CSoprxb-YcJy-YUKiTcYkxN8sb9h2kdk_iHuncvSV5c,9683
 liger_kernel/ops/fused_neighborhood_attention.py,sha256=vPi5xbnh6wxyZehaqo6Tuilqo2fN5SGDiONjnNmIKqs,35556
 liger_kernel/ops/geglu.py,sha256=r0WSq9E93zzynL44Wh8femzOWK07_SseBM_pJUyxT3s,4144
@@ -58,7 +58,7 @@ liger_kernel/transformers/jsd.py,sha256=DGqRnxIZxsvxo0_tbbxX3b-sDbDjC_yKufyRIHCc
 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=L5mq5mL0GC62bxthN7p4Db5l7NogFE-1JsbZsr4GGik,105877
+liger_kernel/transformers/monkey_patch.py,sha256=TUmx8aY0lonyThcATirRBdSs7uItVvnBggohjBItBuQ,106060
 liger_kernel/transformers/multi_token_attention.py,sha256=K3NIY9_5TPgZ4_Rahn0xnkMXxD_fmlJHK4CWGYvGQp0,1752
 liger_kernel/transformers/qwen2vl_mrope.py,sha256=5EwSqrMdsL9MYspeBMXBsNJKvH0MOmRrtJXAJlnnlOI,1047
 liger_kernel/transformers/rms_norm.py,sha256=vkekcvTeWY8vL4H6hg3t0XeY0Ew_3OFMPHuzqlxPPVw,2719
@@ -99,9 +99,9 @@ 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-0.6.2.dev20251011154226.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel_nightly-0.6.2.dev20251011154226.dist-info/METADATA,sha256=Cu1iB8uZyEsoDEJqbJYGCxahIaLwCKhvp3M1uxaV1Uk,24777
-liger_kernel_nightly-0.6.2.dev20251011154226.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel_nightly-0.6.2.dev20251011154226.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
-liger_kernel_nightly-0.6.2.dev20251011154226.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel_nightly-0.6.2.dev20251011154226.dist-info/RECORD,,
+liger_kernel_nightly-0.6.2.dev20251013144132.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel_nightly-0.6.2.dev20251013144132.dist-info/METADATA,sha256=3lZjwj_uIcS1aYE--_B3JuOh95x-txytvJPkdZGO_QA,24777
+liger_kernel_nightly-0.6.2.dev20251013144132.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel_nightly-0.6.2.dev20251013144132.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
+liger_kernel_nightly-0.6.2.dev20251013144132.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel_nightly-0.6.2.dev20251013144132.dist-info/RECORD,,