liger-kernel 0.5.0__py3-none-any.whl → 0.5.2__py3-none-any.whl

liger_kernel/ops/qwen2vl_mrope.py

@@ -10,6 +10,7 @@ def _triton_qwen2vl_mrope(
     cos,
     sin,
     sl,
+    bs: tl.constexpr,
     n_qh: tl.constexpr,
     n_kh: tl.constexpr,
     hd: tl.constexpr,
@@ -41,13 +42,12 @@ def _triton_qwen2vl_mrope(
     t_end = mrope_section_t
     h_end = t_end + mrope_section_h
 
-    cos_row_idx = pid % sl
-    t_cos = cos + cos_row_idx * hd
-    h_cos = t_cos + sl * hd
-    w_cos = h_cos + sl * hd
-    t_sin = sin + cos_row_idx * hd
-    h_sin = t_sin + sl * hd
-    w_sin = h_sin + sl * hd
+    t_cos = cos + pid * hd
+    h_cos = t_cos + bs * sl * hd
+    w_cos = h_cos + bs * sl * hd
+    t_sin = sin + pid * hd
+    h_sin = t_sin + bs * sl * hd
+    w_sin = h_sin + bs * sl * hd
 
     cos_offsets = tl.arange(0, pad_hd // 2)
     t_mask = cos_offsets < t_end
@@ -151,6 +151,7 @@ def qwen2vl_mrope_forward(q, k, cos, sin, mrope_section):
         cos,
         sin,
         seq_len,
+        batch_size,
         n_q_head,
         n_kv_head,
         head_dim,
@@ -189,6 +190,7 @@ def qwen2vl_mrope_backward(dq, dk, cos, sin, mrope_section):
         cos,
         sin,
         seq_len,
+        batch_size,
         n_q_head,
         n_kv_head,
         head_dim,
@@ -216,8 +218,8 @@ class LigerQwen2VLMRopeFunction(torch.autograd.Function):
         """
         q size: (bsz, n_q_head, seq_len, head_dim)
         k size: (bsz, n_kv_head, seq_len, head_dim)
-        cos size: (3, 1, seq_len, head_dim)
-        sin size: (3, 1, seq_len, head_dim)
+        cos size: (3, bsz, seq_len, head_dim)
+        sin size: (3, bsz, seq_len, head_dim)
         """
         q, k, cos, sin = qwen2vl_mrope_forward(q, k, cos, sin, mrope_section)
         ctx.save_for_backward(cos, sin)
@@ -228,10 +230,9 @@ class LigerQwen2VLMRopeFunction(torch.autograd.Function):
         """
         dq size: (bsz, n_q_head, seq_len, head_dim)
         dk size: (bsz, n_kv_head, seq_len, head_dim)
-        cos size: (3, 1, seq_len, head_dim)
-        sin size: (3, 1, seq_len, head_dim)
+        cos size: (3, bsz, seq_len, head_dim)
+        sin size: (3, bsz, seq_len, head_dim)
         """
-
         cos, sin = ctx.saved_tensors
         mrope_section = ctx.mrope_section
         dq, dk = qwen2vl_mrope_backward(dq, dk, cos, sin, mrope_section)

liger_kernel/transformers/__init__.py

@@ -22,7 +22,6 @@ from liger_kernel.transformers.monkey_patch import (  # noqa: F401
     apply_liger_kernel_to_qwen2,
     apply_liger_kernel_to_qwen2_vl,
 )
-from liger_kernel.transformers.orpo_trainer import LigerORPOTrainer  # 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 (  # noqa: F401

liger_kernel/transformers/qwen2vl_mrope.py

@@ -8,8 +8,8 @@ def liger_multimodal_rotary_pos_emb(q, k, cos, sin, mrope_section, unsqueeze_dim
     Args:
         q (torch.Tensor): The query tensor of shape (bsz, n_q_head, seq_len, head_dim).
         k (torch.Tensor): The key tensor of shape (bsz, n_kv_head, seq_len, head_dim).
-        cos (torch.Tensor): The cosine tensor of shape (3, 1, seq_len, head_dim).
-        sin (torch.Tensor): The sine tensor of shape (3, 1, seq_len, head_dim).
+        cos (torch.Tensor): The cosine tensor of shape (3, bsz, seq_len, head_dim).
+        sin (torch.Tensor): The sine tensor of shape (3, bsz, seq_len, head_dim).
         mrope_section (List[int]): The multimodal rope section for channel dimension of temporal, height and width in rope calculation.
         unsqueeze_dim (int, optional): The dimension to unsqueeze. Defaults to 1.
 

liger_kernel/transformers/trainer/__init__.py

@@ -0,0 +1,6 @@
+try:
+    from liger_kernel.transformers.trainer.orpo_trainer import (  # noqa: F401
+        LigerORPOTrainer,
+    )
+except ImportError:
+    raise ImportError("Please `pip install trl` to use LigerORPOTrainer")

liger_kernel/transformers/{orpo_trainer.py → trainer/orpo_trainer.py}

@@ -76,9 +76,7 @@ class LigerORPOTrainer(ORPOTrainer):
             padding_value=self.padding_value,
             device=self.accelerator.device,
         )
-        # if self.accelerator.is_main_process:
-        #     import pdb; pdb.set_trace()
-        # torch.distributed.barrier()
+
         model_kwargs = (
             {
                 "decoder_input_ids": self._shift_right(

{liger_kernel-0.5.0.dist-info → liger_kernel-0.5.2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel
-Version: 0.5.0
+Version: 0.5.2
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -34,9 +34,10 @@ Requires-Dist: torch>=2.1.2
 Requires-Dist: triton>=2.3.1
 Provides-Extra: transformers
 Requires-Dist: transformers~=4.0; extra == "transformers"
+Provides-Extra: trl
+Requires-Dist: trl>=0.11.0; extra == "trl"
 Provides-Extra: dev
 Requires-Dist: transformers>=4.44.2; extra == "dev"
-Requires-Dist: trl>=0.11.0; extra == "dev"
 Requires-Dist: matplotlib>=3.7.2; extra == "dev"
 Requires-Dist: flake8>=4.0.1.1; extra == "dev"
 Requires-Dist: black>=24.4.2; extra == "dev"

{liger_kernel-0.5.0.dist-info → liger_kernel-0.5.2.dist-info}/RECORD

@@ -18,14 +18,14 @@ liger_kernel/ops/group_norm.py,sha256=VaRErVJGR4JqgXXvuIjNGTn3E2egjLtU1y3ymwIf4d
 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=_CZggw3GNEIUx5weDzadFit5I-Lzosoo8prgeJzcViY,7589
-liger_kernel/ops/qwen2vl_mrope.py,sha256=xZvQnhkSTjU-k6KiiRn9e0SYO1ESs1jmuZFMICduLpc,8552
+liger_kernel/ops/qwen2vl_mrope.py,sha256=GvP4Cg-2ClYyiqbe7bB_OMvnlZooBmqP2-9V8RMPde4,8598
 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=_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=P5JR3fI-znhG92nRrFS2j0TIJTLhP-xD5dvEy4HP9ik,1418
+liger_kernel/transformers/__init__.py,sha256=gia-eBxr7TLxU0GdDf8AfCY4WgDlFLqIGSt7EoQGsBA,1336
 liger_kernel/transformers/auto_model.py,sha256=RMIwQHSiXoksXFTIqFZ4PLBgoqkxJJAT3q1Qh47bGN8,1552
 liger_kernel/transformers/cross_entropy.py,sha256=yEm_YQ7oa3_BzT3hdW6KrAslduhSqWcJQVNZZDcWCg4,1758
 liger_kernel/transformers/functional.py,sha256=sUBoU8Vb4pLpr9G6IdkRsToYgh-rCXL4OLYat7Tv_GU,4450
@@ -37,8 +37,7 @@ liger_kernel/transformers/jsd.py,sha256=sbr8DnKSYZJH9pv2rpmboNijYGpZKbhb2-WSGp5_
 liger_kernel/transformers/kl_div.py,sha256=qVhjBg6tjRyue5iZ3NFxo8uySY4JuIFJyv0IM_50F24,431
 liger_kernel/transformers/layer_norm.py,sha256=fd6o4kSHJWolQMWxh-l1qObfgL08ruNbUoBiANKX1ow,972
 liger_kernel/transformers/monkey_patch.py,sha256=Fk2v4GZQDJzfh3Cpc6BHNJbs_tungDyWmqS9nuG9Lc4,38406
-liger_kernel/transformers/orpo_trainer.py,sha256=mC8ePS-Oq-BrdM0lKpgSBLuYLqYsWxH_4Q2RnDthz5M,7643
-liger_kernel/transformers/qwen2vl_mrope.py,sha256=SfSQVwOe7ArrVfpmIdfZrdzCxmcj7V-YQp9zDu17-ao,1043
+liger_kernel/transformers/qwen2vl_mrope.py,sha256=5EwSqrMdsL9MYspeBMXBsNJKvH0MOmRrtJXAJlnnlOI,1047
 liger_kernel/transformers/rms_norm.py,sha256=AHstklNIO1PLHjjCBU-TPuUD-Fl_pycJUTLlJNojbV8,1189
 liger_kernel/transformers/rope.py,sha256=m-ah8vZBYW8tfplTXCiAPMHJWlB1tdp_JPXJeWE-Boo,943
 liger_kernel/transformers/swiglu.py,sha256=0-tVJ8xEYfhxnduc16PflXFj8sZPxdx9sHUn3hfwCI4,2468
@@ -54,11 +53,13 @@ liger_kernel/transformers/model/mllama.py,sha256=mesNCgj0Ea1O-fqRD4LVxDJ1CR2abY_
 liger_kernel/transformers/model/phi3.py,sha256=xUZPlaPKwknLjHc3uUW3EPodm1h0vD3G7Qnhh51v-Io,10332
 liger_kernel/transformers/model/qwen2.py,sha256=EyhSSzQOskGjSnCsKMZpd1s5IAIlHd5PBO3q0MoCs00,9619
 liger_kernel/transformers/model/qwen2_vl.py,sha256=bIQe2bWiY--G84FhCD29Gdi64_qHP6vbcGsK6vKysQE,8547
+liger_kernel/transformers/trainer/__init__.py,sha256=c4OQVJmhNOloj0JYSEc0j_cQuBbzGWILfaowUR1hmRw,210
+liger_kernel/transformers/trainer/orpo_trainer.py,sha256=jko6oq_XQdBSmXubp05E-_YXOyhtB5Bj75dg5YNwOsE,7517
 liger_kernel/triton/__init__.py,sha256=yfRe0zMb47QnqjecZWG7LnanfCTzeku7SgWRAwNVmzU,101
 liger_kernel/triton/monkey_patch.py,sha256=5BcGKTtdqeYchypBIBopGIWPx1-cFALz7sOKoEsqXJ0,1584
-liger_kernel-0.5.0.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel-0.5.0.dist-info/METADATA,sha256=7c5Tzf84zfQFOdXxx5nXg0wqGKH8VhsLCfTvoMN3kNM,20675
-liger_kernel-0.5.0.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel-0.5.0.dist-info/WHEEL,sha256=PZUExdf71Ui_so67QXpySuHtCi3-J3wvF4ORK6k_S8U,91
-liger_kernel-0.5.0.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel-0.5.0.dist-info/RECORD,,
+liger_kernel-0.5.2.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel-0.5.2.dist-info/METADATA,sha256=olSIT-Jd2Mowu2ja4QLwyPYBhCnY22znBq9pV7stkKI,20695
+liger_kernel-0.5.2.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel-0.5.2.dist-info/WHEEL,sha256=PZUExdf71Ui_so67QXpySuHtCi3-J3wvF4ORK6k_S8U,91
+liger_kernel-0.5.2.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel-0.5.2.dist-info/RECORD,,