liger-kernel-nightly 0.4.2.dev20241119054334__py3-none-any.whl → 0.4.2.dev20241119054537__py3-none-any.whl

liger_kernel/ops/qwen2vl_mrope.py

@@ -0,0 +1,238 @@
+import torch
+import triton
+import triton.language as tl
+
+
+@triton.jit
+def _triton_qwen2vl_mrope(
+    q_ptr,
+    k_ptr,
+    cos,
+    sin,
+    sl,
+    n_qh: tl.constexpr,
+    n_kh: tl.constexpr,
+    hd: tl.constexpr,
+    pad_n_qh: tl.constexpr,
+    pad_n_kh: tl.constexpr,
+    pad_hd: tl.constexpr,
+    mrope_section_t: tl.constexpr,
+    mrope_section_h: tl.constexpr,
+    BLOCK_SIZE: tl.constexpr,
+    BACKWARD_PASS: tl.constexpr = False,
+):
+    pid = tl.program_id(0)
+
+    # locate start address
+    q_ptr = q_ptr + pid * (n_qh * hd)
+    k_ptr = k_ptr + pid * (n_kh * hd)
+
+    # ####################################################################
+    # get the cos(mθ_{i...d/2}) and sin(mθ_{i...d/2}) for token position
+    # m of this program instance
+    # ####################################################################
+
+    # 1. program instances are laid out in a 1D vector of size bsz * seq_len, which
+    # effectively represents a 2D grid of size [bsz, seq_len] with seq_len dimension
+    # being the fastest changing dimension. Thus we can simply do pid // sl to get the batch index
+    # and pid % sl to get the sequence index.
+    # 2. We only need the left half of cos and sin matrix because the right half is just
+    # a clone of the left half.
+    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
+
+    cos_offsets = tl.arange(0, pad_hd // 2)
+    t_mask = cos_offsets < t_end
+    h_mask = (t_end <= cos_offsets) & (cos_offsets < h_end)
+    w_mask = (h_end <= cos_offsets) & (cos_offsets < hd // 2)
+    t_cos_row = tl.load(t_cos + cos_offsets, mask=t_mask, other=0)
+    h_cos_row = tl.load(h_cos + cos_offsets, mask=h_mask, other=0)
+    w_cos_row = tl.load(w_cos + cos_offsets, mask=w_mask, other=0)
+    t_sin_row = tl.load(t_sin + cos_offsets, mask=t_mask, other=0)
+    h_sin_row = tl.load(h_sin + cos_offsets, mask=h_mask, other=0)
+    w_sin_row = tl.load(w_sin + cos_offsets, mask=w_mask, other=0)
+    cos_row = t_cos_row + h_cos_row + w_cos_row
+    sin_row = t_sin_row + h_sin_row + w_sin_row
+
+    # ####################################################################
+    # Load the left and right half of q and k for the current
+    # program instance (i.e. for the current token) separately
+    # ####################################################################
+    # left half of the head
+    first_half_q_offsets = (
+        tl.arange(0, pad_n_qh)[:, None] * hd + tl.arange(0, pad_hd // 2)[None, :]
+    )
+    first_half_k_offsets = (
+        tl.arange(0, pad_n_kh)[:, None] * hd + tl.arange(0, pad_hd // 2)[None, :]
+    )
+    first_q_mask = (tl.arange(0, pad_n_qh)[:, None] < n_qh) & (
+        tl.arange(0, pad_hd // 2)[None, :] < hd // 2
+    )
+    first_k_mask = (tl.arange(0, pad_n_kh)[:, None] < n_kh) & (
+        tl.arange(0, pad_hd // 2)[None, :] < hd // 2
+    )
+    q_tile_1 = tl.load(q_ptr + first_half_q_offsets, mask=first_q_mask, other=0).to(
+        sin_row.dtype
+    )
+    k_tile_1 = tl.load(k_ptr + first_half_k_offsets, mask=first_k_mask, other=0).to(
+        sin_row.dtype
+    )
+
+    # right half of the head
+    second_half_q_offsets = first_half_q_offsets + (hd // 2)
+    second_half_k_offsets = first_half_k_offsets + (hd // 2)
+    second_q_mask = first_q_mask
+    second_k_mask = first_k_mask
+    q_tile_2 = tl.load(q_ptr + second_half_q_offsets, mask=second_q_mask, other=0).to(
+        sin_row.dtype
+    )
+    k_tile_2 = tl.load(k_ptr + second_half_k_offsets, mask=second_k_mask, other=0).to(
+        sin_row.dtype
+    )
+
+    if not BACKWARD_PASS:
+        # y = [x1, x2] * [cos, cos] + [-x2, x1] * [sin, sin]
+        new_q_tile_1 = q_tile_1 * cos_row - q_tile_2 * sin_row
+        tl.store(q_ptr + first_half_q_offsets, new_q_tile_1, mask=first_q_mask)
+        new_q_tile_2 = q_tile_2 * cos_row + q_tile_1 * sin_row
+        tl.store(q_ptr + second_half_q_offsets, new_q_tile_2, mask=second_q_mask)
+
+        new_k_tile_1 = k_tile_1 * cos_row - k_tile_2 * sin_row
+        tl.store(k_ptr + first_half_k_offsets, new_k_tile_1, mask=first_k_mask)
+        new_k_tile_2 = k_tile_2 * cos_row + k_tile_1 * sin_row
+        tl.store(k_ptr + second_half_k_offsets, new_k_tile_2, mask=second_k_mask)
+    else:
+        # with some math, we can get:
+        # dy = [dx1, dx2] * [cos, cos] + [-dx2, dx1] * [-sin, -sin]
+        new_q_tile_1 = q_tile_1 * cos_row + q_tile_2 * sin_row
+        tl.store(q_ptr + first_half_q_offsets, new_q_tile_1, mask=first_q_mask)
+        new_q_tile_2 = q_tile_2 * cos_row - q_tile_1 * sin_row
+        tl.store(q_ptr + second_half_q_offsets, new_q_tile_2, mask=second_q_mask)
+
+        new_k_tile_1 = k_tile_1 * cos_row + k_tile_2 * sin_row
+        tl.store(k_ptr + first_half_k_offsets, new_k_tile_1, mask=first_k_mask)
+        new_k_tile_2 = k_tile_2 * cos_row - k_tile_1 * sin_row
+        tl.store(k_ptr + second_half_k_offsets, new_k_tile_2, mask=second_k_mask)
+
+
+def qwen2vl_mrope_forward(q, k, cos, sin, mrope_section):
+
+    # transpose it back to the physical shape because Triton looks at the physical storage
+    # note: q and k are incontiguous before the transformation and will become contiguous after transpose
+    q = q.transpose(1, 2)
+    k = k.transpose(1, 2)
+
+    batch_size, seq_len, n_q_head, head_dim = q.shape
+    n_kv_head = k.shape[2]
+    pad_hd = triton.next_power_of_2(head_dim)
+    pad_n_q_head = triton.next_power_of_2(n_q_head)
+    pad_n_kv_head = triton.next_power_of_2(n_kv_head)
+    BLOCK_SIZE = max(pad_n_q_head, pad_n_kv_head)
+
+    n_row = batch_size * seq_len
+
+    # ensure tensors passed into the kernel are contiguous. It will be no-op if they are already contiguous
+    q = q.contiguous()
+    k = k.contiguous()
+    cos = cos.contiguous()
+    sin = sin.contiguous()
+
+    _triton_qwen2vl_mrope[(n_row,)](
+        q,
+        k,
+        cos,
+        sin,
+        seq_len,
+        n_q_head,
+        n_kv_head,
+        head_dim,
+        pad_n_q_head,
+        pad_n_kv_head,
+        pad_hd,
+        mrope_section[0],
+        mrope_section[1],
+        BLOCK_SIZE=BLOCK_SIZE,
+        BACKWARD_PASS=False,
+    )
+    return q.transpose(1, 2), k.transpose(1, 2), cos, sin
+
+
+def qwen2vl_mrope_backward(dq, dk, cos, sin, mrope_section):
+    dq = dq.transpose(1, 2)
+    dk = dk.transpose(1, 2)
+
+    batch_size, seq_len, n_q_head, head_dim = dq.shape
+    n_kv_head = dk.shape[2]
+    pad_hd = triton.next_power_of_2(head_dim)
+    pad_n_q_head = triton.next_power_of_2(n_q_head)
+    pad_n_kv_head = triton.next_power_of_2(n_kv_head)
+    BLOCK_SIZE = max(pad_n_q_head, pad_n_kv_head)
+
+    n_row = batch_size * seq_len
+
+    # ensure dq and dk are contiguous
+    dq = dq.contiguous()
+    dk = dk.contiguous()
+
+    # backward is similar to forward except swapping few ops
+    _triton_qwen2vl_mrope[(n_row,)](
+        dq,
+        dk,
+        cos,
+        sin,
+        seq_len,
+        n_q_head,
+        n_kv_head,
+        head_dim,
+        pad_n_q_head,
+        pad_n_kv_head,
+        pad_hd,
+        mrope_section[0],
+        mrope_section[1],
+        BLOCK_SIZE=BLOCK_SIZE,
+        BACKWARD_PASS=True,
+    )
+    return dq.transpose(1, 2), dk.transpose(1, 2)
+
+
+class LigerQwen2VLMRopeFunction(torch.autograd.Function):
+    """
+    Triton implementation of the Qwen2VL Multimodal Rotary Positional Embedding (M-RoPE) operation.
+
+    Please find the corresponding HuggingFace implementation here:
+    https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen2_vl/modeling_qwen2_vl.py
+    """
+
+    @staticmethod
+    def forward(ctx, q, k, cos, sin, mrope_section, unsqueeze_dim=1):
+        """
+        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)
+        """
+        q, k, cos, sin = qwen2vl_mrope_forward(q, k, cos, sin, mrope_section)
+        ctx.save_for_backward(cos, sin)
+        ctx.mrope_section = mrope_section
+        return q, k
+
+    def backward(ctx, dq, dk):
+        """
+        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, sin = ctx.saved_tensors
+        mrope_section = ctx.mrope_section
+        dq, dk = qwen2vl_mrope_backward(dq, dk, cos, sin, mrope_section)
+        return dq, dk, None, None, None, None

liger_kernel/transformers/functional.py

@@ -10,6 +10,7 @@ from liger_kernel.ops.group_norm import LigerGroupNormFunction
 from liger_kernel.ops.jsd import LigerJSDFunction
 from liger_kernel.ops.kl_div import LigerKLDivLossFunction
 from liger_kernel.ops.layer_norm import LigerLayerNormFunction
+from liger_kernel.ops.qwen2vl_mrope import LigerQwen2VLMRopeFunction
 from liger_kernel.ops.rms_norm import LigerRMSNormFunction
 from liger_kernel.ops.rope import LigerRopeFunction
 from liger_kernel.ops.swiglu import LigerSiLUMulFunction
@@ -19,6 +20,7 @@ liger_fused_linear_cross_entropy = LigerFusedLinearCrossEntropyFunction.apply
 liger_geglu = LigerGELUMulFunction.apply
 liger_rms_norm = LigerRMSNormFunction.apply
 liger_rope = LigerRopeFunction.apply
+liger_qwen2vl_mrope = LigerQwen2VLMRopeFunction.apply
 liger_layer_norm = LigerLayerNormFunction.apply
 liger_kl_div = LigerKLDivLossFunction.apply
 liger_jsd = LigerJSDFunction.apply

liger_kernel/transformers/monkey_patch.py

@@ -36,6 +36,7 @@ from liger_kernel.transformers.model.qwen2 import lce_forward as qwen2_lce_forwa
 from liger_kernel.transformers.model.qwen2 import (
     lce_forward_deprecated as qwen2_lce_forward_deprecated,
 )
+from liger_kernel.transformers.qwen2vl_mrope import liger_multimodal_rotary_pos_emb
 from liger_kernel.transformers.rms_norm import LigerRMSNorm
 from liger_kernel.transformers.rope import liger_rotary_pos_emb
 from liger_kernel.transformers.swiglu import (
@@ -610,9 +611,7 @@ def apply_liger_kernel_to_qwen2(
             logger.warning(TRANSFORMER_DEPRECATION_WARNING)
             modeling_qwen2.CrossEntropyLoss = LigerCrossEntropyLoss
 
-    # import pdb; pdb.set_trace()
     if fused_linear_cross_entropy:
-
         if transformer_version >= version.parse(SUPPORTED_TRANSFORMER_VERSION):
             modeling_qwen2.Qwen2ForCausalLM.forward = qwen2_lce_forward
         else:  # if version < 4.46.1
@@ -644,6 +643,7 @@ def apply_liger_kernel_to_qwen2(
 
 
 def apply_liger_kernel_to_qwen2_vl(
+    rope: bool = True,
     cross_entropy: bool = False,
     fused_linear_cross_entropy: bool = True,
     rms_norm: bool = True,
@@ -678,8 +678,10 @@ def apply_liger_kernel_to_qwen2_vl(
         lce_forward as qwen2_vl_lce_forward,
     )
 
-    # TODO: Support Qwen2-VL's multimodal RoPE implementation
-
+    if rope:
+        modeling_qwen2_vl.apply_multimodal_rotary_pos_emb = (
+            liger_multimodal_rotary_pos_emb
+        )
     if rms_norm:
         # https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen2_vl/modeling_qwen2_vl.py#L439
         modeling_qwen2_vl.Qwen2RMSNorm = LigerRMSNorm

liger_kernel/transformers/qwen2vl_mrope.py

@@ -0,0 +1,20 @@
+from liger_kernel.ops.qwen2vl_mrope import LigerQwen2VLMRopeFunction
+
+
+def liger_multimodal_rotary_pos_emb(q, k, cos, sin, mrope_section, unsqueeze_dim=1):
+    """
+    Applies Multimodal Rotary Positional Embedding (M-RoPE) operation to query and key states.
+
+    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).
+        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.
+
+    Returns:
+        Tuple[torch.Tensor, torch.Tensor]: The query and key tensors after applying the M-RoPE operation.
+    """
+
+    return LigerQwen2VLMRopeFunction.apply(q, k, cos, sin, mrope_section, unsqueeze_dim)

{liger_kernel_nightly-0.4.2.dev20241119054334.dist-info → liger_kernel_nightly-0.4.2.dev20241119054537.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.4.2.dev20241119054334
+Version: 0.4.2.dev20241119054537
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.4.2.dev20241119054334.dist-info → liger_kernel_nightly-0.4.2.dev20241119054537.dist-info}/RECORD

@@ -13,6 +13,7 @@ liger_kernel/ops/group_norm.py,sha256=VaRErVJGR4JqgXXvuIjNGTn3E2egjLtU1y3ymwIf4d
 liger_kernel/ops/jsd.py,sha256=anWfdioucxZy4JQfTvbHBR-IQrZKeH-gBF1MHwwTuTQ,5781
 liger_kernel/ops/kl_div.py,sha256=03FNXfvCb6M-56hhFepAFV9p6brArPR6KOKkdGD34mw,8374
 liger_kernel/ops/layer_norm.py,sha256=unGMYMOPqtkM9aTrokhcqgPmsV2AUN7Yzv86isVB9OI,7422
+liger_kernel/ops/qwen2vl_mrope.py,sha256=xZvQnhkSTjU-k6KiiRn9e0SYO1ESs1jmuZFMICduLpc,8552
 liger_kernel/ops/rms_norm.py,sha256=LAxCiFjpBbb7TDh9pOzsVmDGAR7eEbTDnEhjSd6TX_M,11583
 liger_kernel/ops/rope.py,sha256=jrzaA9-6Orn44y_IIam9_YNPQxOFK2FrIRNfFea4EtU,8513
 liger_kernel/ops/swiglu.py,sha256=Fwxtd76rhHKT9ShQAGca9RsnASplAVxtYKHmiT73_yA,2994
@@ -22,7 +23,7 @@ liger_kernel/ops/experimental/mm_int8int2.py,sha256=JpGVZCgRC6T8XMUJ_QbZRS2XU1bh
 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=Hd4WvxNqOJHM9HmRfAQueRnmOy5WU9nFsFygB5Iv8Xs,2000
+liger_kernel/transformers/functional.py,sha256=jwTHmyjOVC1_I-6ztY1EbbRqPIfFHojcHrP2c4P6U4I,2123
 liger_kernel/transformers/fused_linear_cross_entropy.py,sha256=_i0PXSp5iZ9pKXdEeZ4lvHCENJYjV4y74yz3ZRG5XQg,1484
 liger_kernel/transformers/fused_linear_jsd.py,sha256=MJ-KjmLZnakuoVpnbDGkd95DQgvESniyrRWYzollVZM,4066
 liger_kernel/transformers/geglu.py,sha256=QcrME_8ooIn0xa59LaC0aoOdRrBIFd11Y0bAyF0NfCw,1130
@@ -30,7 +31,8 @@ liger_kernel/transformers/group_norm.py,sha256=FJ9R7mS9G1wO-GRIQ6QKSmIhnZ6nQ6GIk
 liger_kernel/transformers/jsd.py,sha256=W-5CypO2mx4-bUWOxq1KScfCdoXlLoYbtt5xBnRzMs4,3056
 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=L1IuGmFMWYgf-u3OXCg43BUxbZKTpd7ATjjDjYoFkEM,38268
+liger_kernel/transformers/monkey_patch.py,sha256=Fk2v4GZQDJzfh3Cpc6BHNJbs_tungDyWmqS9nuG9Lc4,38406
+liger_kernel/transformers/qwen2vl_mrope.py,sha256=SfSQVwOe7ArrVfpmIdfZrdzCxmcj7V-YQp9zDu17-ao,1043
 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
@@ -48,9 +50,9 @@ 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.dev20241119054334.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel_nightly-0.4.2.dev20241119054334.dist-info/METADATA,sha256=xMe_8AJmITw8CPUUvlnKZT2vwaL9doaM6l8LdLwscpI,21556
-liger_kernel_nightly-0.4.2.dev20241119054334.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel_nightly-0.4.2.dev20241119054334.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
-liger_kernel_nightly-0.4.2.dev20241119054334.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel_nightly-0.4.2.dev20241119054334.dist-info/RECORD,,
+liger_kernel_nightly-0.4.2.dev20241119054537.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel_nightly-0.4.2.dev20241119054537.dist-info/METADATA,sha256=BoWJXhq2CldcpNJomRA4lKZ4J2AEr9tVfrjGyxq3EdM,21556
+liger_kernel_nightly-0.4.2.dev20241119054537.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel_nightly-0.4.2.dev20241119054537.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
+liger_kernel_nightly-0.4.2.dev20241119054537.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel_nightly-0.4.2.dev20241119054537.dist-info/RECORD,,