liger-kernel-nightly 0.4.2.dev20241122052539__py3-none-any.whl → 0.4.2.dev20241123040418__py3-none-any.whl

liger_kernel/ops/cross_entropy.py

@@ -92,8 +92,8 @@ def liger_cross_entropy_kernel(
     # 3. [Online softmax] first pass: find max + sum
     m = float("-inf")  # m is the max value. use the notation from the paper
     d = 0.0  # d is the sum. use the notation from the paper
-    ori_X_y = tl.load(
-        X_ptr + y
+    ori_X_y = tl.load(X_ptr + y).cast(
+        tl.float32
     )  # we need to store the original value of X_y for the loss calculation
     if HAS_SOFTCAPPING:
         ori_X_y = softcap * tanh(ori_X_y / softcap)
@@ -106,8 +106,11 @@ def liger_cross_entropy_kernel(
     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")
-        )
+            X_ptr + X_offsets,
+            mask=X_offsets < n_cols,
+            other=float("-inf"),
+            # Ensure float32 precision for softmax calculation
+        ).cast(tl.float32)
         if HAS_SOFTCAPPING:
             X_block = softcap * tanh(X_block / softcap)
         block_max = tl.max(X_block)
@@ -141,8 +144,11 @@ def liger_cross_entropy_kernel(
     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")
-        )
+            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

liger_kernel/ops/fused_linear_cross_entropy.py

@@ -26,7 +26,6 @@ def fused_linear_cross_entropy_forward(
     reduction="mean",
     softcap=None,
 ):
-    dtype = _input.dtype
     device = _input.device
 
     # inputs have shape: BT x H
@@ -74,9 +73,6 @@ def fused_linear_cross_entropy_forward(
         loss_1d_slice = loss_1d[start_idx:end_idx]  # chunk_size,
         n_non_ignore = (target_chunk != ignore_index).sum().item()
 
-        # when doing CE, use the upcasted precision
-        logits_chunk = logits_chunk.float()
-
         # ensure _input and target are contiguous
         logits_chunk = logits_chunk.contiguous()
         target_chunk = target_chunk.contiguous()
@@ -103,13 +99,6 @@ def fused_linear_cross_entropy_forward(
             num_warps=32 if not is_hip() else 16,
         )
 
-        # gradient of logits_chunk is computed in-place by the above triton kernel.
-        # Following HuggingFace model source code, we do the forward and backward
-        # w.r.t. logits in fp32 for numerical stability especially as the num classes (vocab size) is huge.
-        # (reference: https://github.com/huggingface/transformers/blob/v4.42.4/src/transformers/models/llama/modeling_llama.py#L1194)
-        # Propagating to lm_head's backward, we'll switch back to the original dtype.
-        logits_chunk = logits_chunk.to(dtype)
-
         # gradient of logits_chunk is computed in-place by the above triton kernel and is of shape: chunk_size x V
         # thus grad_input[start_idx: end_idx] should be of shape: chunk_size x H
         # additionally, since we are chunking the inputs, observe that the loss and gradients are calculated only

liger_kernel/ops/layer_norm.py

@@ -180,8 +180,13 @@ def layer_norm_backward(dY, X, W, B, Mean, RSTD):
     dY = dY.view(-1, dim)
     n_rows, n_cols = dY.shape
 
+    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_subslice_count
+
     DX = torch.empty((n_rows, n_cols), dtype=X.dtype, device=X.device)
-    sm_count = torch.cuda.get_device_properties(X.device).multi_processor_count
     _DW = torch.empty((sm_count, n_cols), dtype=W.dtype, device=W.device)
     _DB = torch.empty((sm_count, n_cols), dtype=W.dtype, device=W.device)
 

liger_kernel/ops/rms_norm.py

@@ -264,6 +264,7 @@ def rms_norm_backward(
     dY = dY.view(-1, dim)
     n_rows, n_cols = dY.shape
 
+    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":

liger_kernel/ops/utils.py

@@ -20,6 +20,8 @@ import triton
 import triton.language as tl
 from packaging.version import Version
 
+from liger_kernel.utils import infer_device
+
 
 def is_hip() -> bool:
     return torch.version.hip is not None
@@ -69,10 +71,11 @@ def compare_version(package: str, operator: Callable, target: str):
 
 
 def get_amp_custom_fwd_bwd() -> Callable:
+    device = infer_device()
     if compare_version("torch", operator.ge, "2.4.0"):
         return (
-            functools.partial(torch.amp.custom_fwd, device_type="cuda"),
-            functools.partial(torch.amp.custom_bwd, device_type="cuda"),
+            functools.partial(torch.amp.custom_fwd, device_type=device),
+            functools.partial(torch.amp.custom_bwd, device_type=device),
         )
     return torch.cuda.amp.custom_fwd, torch.cuda.amp.custom_bwd
 

liger_kernel/utils.py

@@ -0,0 +1,13 @@
+import torch
+
+
+def infer_device():
+    """
+    Get current device name based on available devices
+    """
+    if torch.cuda.is_available():
+        return "cuda"
+    elif torch.xpu.is_available():
+        return "xpu"
+    else:
+        return "cpu"

{liger_kernel_nightly-0.4.2.dev20241122052539.dist-info → liger_kernel_nightly-0.4.2.dev20241123040418.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.4.2.dev20241122052539
+Version: 0.4.2.dev20241123040418
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.4.2.dev20241122052539.dist-info → liger_kernel_nightly-0.4.2.dev20241123040418.dist-info}/RECORD

@@ -1,4 +1,6 @@
+liger_kernel/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 liger_kernel/env_report.py,sha256=jye8RvUkmhqaIshdeIpoUABoAu7FPKJUib4FnAfvkpw,1132
+liger_kernel/utils.py,sha256=HJa-xVKOohDn6pLVIx-Fv0V9h0QAL3qZGQNRICI-OpI,249
 liger_kernel/chunked_loss/__init__.py,sha256=R2wCcz4Y0kTAve926DH3k182XKezpXeACMHj05g9Mm8,346
 liger_kernel/chunked_loss/cpo_loss.py,sha256=H2L6mNtU8RMJ17u4aMZ9FHEfBvg1Z_hliY5-jZxiDBM,3079
 liger_kernel/chunked_loss/dpo_loss.py,sha256=XcCGLVmTVdEX30q41XRXXK_c-MSumVJ-l4tQwobUv2w,4228
@@ -7,19 +9,19 @@ liger_kernel/chunked_loss/fused_linear_preference.py,sha256=nkEpNWTHh5GmlnHOnGx5
 liger_kernel/chunked_loss/orpo_loss.py,sha256=DZ-_hm1twllBWujEV4M4-VDBkxMDBvoGqMGe-aGP1hA,3147
 liger_kernel/chunked_loss/simpo_loss.py,sha256=Jpl_U6DfxlzyHnlKN2i05K0vwz-ouiTmxlLGb439FwY,3328
 liger_kernel/ops/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-liger_kernel/ops/cross_entropy.py,sha256=sfUb7-jIZp0EKXjg1DYy2Wdzw_Mg-mHmGoR5bpdm4tw,15526
-liger_kernel/ops/fused_linear_cross_entropy.py,sha256=ib7M3AjJE164yMfuS9R39k-5qnDgYOXptIT146lqYbg,9964
+liger_kernel/ops/cross_entropy.py,sha256=VqaYB9Zirc51eZ28OmjEZRrrV9UysRjS_vhIftB9sKo,15753
+liger_kernel/ops/fused_linear_cross_entropy.py,sha256=Tnw4gyAYVVdnCOqhOuLEzbUQ3goOTnoAfk3pqSIM5ac,9301
 liger_kernel/ops/fused_linear_jsd.py,sha256=nOv4zwfxHqqepKEmMsQuz-B3H-gRjyo8uClpmqSGLYA,9693
 liger_kernel/ops/geglu.py,sha256=MQL4zyzneZqZYUGPvb1QjI_EYT9_pKfSDgR25WD9jrI,4127
 liger_kernel/ops/group_norm.py,sha256=VaRErVJGR4JqgXXvuIjNGTn3E2egjLtU1y3ymwIf4d8,10961
 liger_kernel/ops/jsd.py,sha256=Ap2b0_geCl6fqBXLI1IS6Yn6GlO-8LgPmnOW3y47dus,6151
 liger_kernel/ops/kl_div.py,sha256=03FNXfvCb6M-56hhFepAFV9p6brArPR6KOKkdGD34mw,8374
-liger_kernel/ops/layer_norm.py,sha256=unGMYMOPqtkM9aTrokhcqgPmsV2AUN7Yzv86isVB9OI,7422
+liger_kernel/ops/layer_norm.py,sha256=_CZggw3GNEIUx5weDzadFit5I-Lzosoo8prgeJzcViY,7589
 liger_kernel/ops/qwen2vl_mrope.py,sha256=xZvQnhkSTjU-k6KiiRn9e0SYO1ESs1jmuZFMICduLpc,8552
-liger_kernel/ops/rms_norm.py,sha256=GKs49wXmUngY7MJ5QDQxTp4P2HDVqzZBaCr0pGtyZyM,11733
+liger_kernel/ops/rms_norm.py,sha256=g7OXwuYI8-LXudDwvXuiupVjjOsbu8c4wwv83VaHa54,11750
 liger_kernel/ops/rope.py,sha256=jrzaA9-6Orn44y_IIam9_YNPQxOFK2FrIRNfFea4EtU,8513
 liger_kernel/ops/swiglu.py,sha256=Fwxtd76rhHKT9ShQAGca9RsnASplAVxtYKHmiT73_yA,2994
-liger_kernel/ops/utils.py,sha256=3JSF--O7KT5Wa5BuO70M4h0XetxoZ_e9IoW9GRlxlBg,3777
+liger_kernel/ops/utils.py,sha256=_VQvd1PX5JXm5xaiBrk2gANp3qr4kM7qYG3ypkBwkMs,3850
 liger_kernel/ops/experimental/embedding.py,sha256=LYR66dB-jhvhtUjeV4PnNro-n77J1mdlmpSLSxB3Y6U,4186
 liger_kernel/ops/experimental/mm_int8int2.py,sha256=JpGVZCgRC6T8XMUJ_QbZRS2XU1bh0urIZphs5DTc1mY,13358
 liger_kernel/transformers/__init__.py,sha256=gia-eBxr7TLxU0GdDf8AfCY4WgDlFLqIGSt7EoQGsBA,1336
@@ -52,9 +54,9 @@ liger_kernel/transformers/model/qwen2.py,sha256=EyhSSzQOskGjSnCsKMZpd1s5IAIlHd5P
 liger_kernel/transformers/model/qwen2_vl.py,sha256=bIQe2bWiY--G84FhCD29Gdi64_qHP6vbcGsK6vKysQE,8547
 liger_kernel/triton/__init__.py,sha256=yfRe0zMb47QnqjecZWG7LnanfCTzeku7SgWRAwNVmzU,101
 liger_kernel/triton/monkey_patch.py,sha256=5BcGKTtdqeYchypBIBopGIWPx1-cFALz7sOKoEsqXJ0,1584
-liger_kernel_nightly-0.4.2.dev20241122052539.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel_nightly-0.4.2.dev20241122052539.dist-info/METADATA,sha256=4fPpoJL_3aqEC-teWIauvl_VbNer1TYZ-7s8znCLSgs,21891
-liger_kernel_nightly-0.4.2.dev20241122052539.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel_nightly-0.4.2.dev20241122052539.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
-liger_kernel_nightly-0.4.2.dev20241122052539.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel_nightly-0.4.2.dev20241122052539.dist-info/RECORD,,
+liger_kernel_nightly-0.4.2.dev20241123040418.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel_nightly-0.4.2.dev20241123040418.dist-info/METADATA,sha256=lXW5-kGkMAutfiUrZflzYeW1bZo8efp65MDohQ3G1T0,21891
+liger_kernel_nightly-0.4.2.dev20241123040418.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel_nightly-0.4.2.dev20241123040418.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
+liger_kernel_nightly-0.4.2.dev20241123040418.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel_nightly-0.4.2.dev20241123040418.dist-info/RECORD,,