rwkv-ops 0.1.0__py3-none-any.whl → 0.1.1__py3-none-any.whl

rwkv_ops/__init__.py

@@ -1,3 +1,4 @@
+__version__ = "0.1.1"
 import os
 
 KERNEL_TYPE = os.environ.get("KERNEL_TYPE", "triton")

rwkv_ops/rwkv7_kernel/__init__.py

@@ -1,5 +1,6 @@
 import keras
 from distutils.util import strtobool
+import os
 from keras import ops
 
 
@@ -19,11 +20,11 @@ def get_generalized_delta_rule(HEAD_SIZE=64, KERNEL_TYPE="native"):
             from .torch_op import generalized_delta_rule
 
             USE_KERNEL = True
+
         elif KERNEL_TYPE.lower() == "cuda":
             CHUNK_LEN = 16
             USE_KERNEL = True
             from torch.utils.cpp_extension import load
-            import os
 
             flags = [
                 "-res-usage",
@@ -137,11 +138,17 @@ def get_generalized_delta_rule(HEAD_SIZE=64, KERNEL_TYPE="native"):
         from jax.lib import xla_bridge
         import jax
         import os
+        import logging
 
+        logging.basicConfig(level=logging.ERROR)
+        os.environ["TRITON_LOG_LEVEL"] = "ERROR"  # 只显示错误级别的日志
+        os.environ["TRITON_DISABLE_AUTOTUNE"] = "1"  # 禁用自动调优日志
+        os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"  # 禁用自动调优日志
         if (
             xla_bridge.get_backend().platform == "gpu"
             and KERNEL_TYPE.lower() == "triton"
         ):
+            os.environ["JAX_LOG_COMPUTATION"] = "0"
             from .jax_op import generalized_delta_rule
 
             USE_KERNEL = True

rwkv_ops/rwkv7_kernel/get_jax_devices_info.py

@@ -5,6 +5,8 @@ import functools
 import triton
 import jax
 import jax.numpy as jnp
+from enum import Enum
+import contextlib
 
 
 @lru_cache(maxsize=None)
@@ -82,9 +84,6 @@ def is_triton_shared_mem_enough(
 
 device_capacity = is_triton_shared_mem_enough()
 
-from enum import Enum
-import contextlib
-
 
 def _cpu_device_warning():
     import warnings

rwkv_ops/rwkv7_kernel/get_torch_devices_info.py

@@ -6,6 +6,8 @@ from typing import Literal
 import triton
 from packaging import version
 import torch
+from enum import Enum
+import contextlib
 
 
 @lru_cache(maxsize=None)
@@ -105,8 +107,6 @@ def is_triton_shared_mem_enough(
 
 
 device_capacity = is_triton_shared_mem_enough()
-from enum import Enum
-import contextlib
 
 
 def _cpu_device_warning():

rwkv_ops/rwkv7_kernel/jax_kernel/chunk_o_bwd.py

@@ -1,5 +1,5 @@
 # -*- coding: utf-8 -*-
-# Copyright (c) 2023-2025,Qingwen Lin
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
 
 from typing import Tuple
 
@@ -7,7 +7,7 @@ import jax_triton as jt
 import jax
 import triton
 
-from ..get_torch_devices_info import check_shared_mem
+from ..get_jax_devices_info import check_shared_mem
 from ..triton_kernel.chunk_o_bwd import *
 
 
@@ -104,7 +104,7 @@ def chunk_dplr_bwd_o(
         out_shape=out_shapes,
         grid=grid,
     )
-    return dq, dk, dw, db, dgk_last
+    return (dq, dk, dw, db, dgk_last)
 
 
 def chunk_dplr_bwd_dAu(

rwkv_ops/rwkv7_kernel/jax_op.py

@@ -223,7 +223,6 @@ def chunk_dplr_bwd(
 
     dv = chunk_dplr_bwd_dv(A_qk=A_qk, kg=kg, do=do, dh=dh, chunk_size=BT)
     del A_qk
-
     dqg, dkg, dw, dbg, dgk_last = chunk_dplr_bwd_o(
         k=kg,
         b=bg,
@@ -239,7 +238,6 @@ def chunk_dplr_bwd(
         scale=scale,
     )
     del v_new
-
     dA_ab, dA_ak, dv, dag = chunk_dplr_bwd_wy(
         A_ab_inv=A_ab_inv,
         A_ak=A_ak,

rwkv_ops/rwkv7_kernel/torch_kernel/chunk_o_bwd.py

@@ -104,7 +104,7 @@ def chunk_dplr_bwd_o(
         BK=BK,
         BV=BV,
     )
-    return dq, dk, dw, db, dgk_last
+    return (dq, dk, dw, db, dgk_last)
 
 
 def chunk_dplr_bwd_dAu(

rwkv_ops/rwkv7_kernel/torch_op.py

@@ -1,25 +1,14 @@
-# -*- coding: utf-8 -*-
-"""
-This file implements the forward and backward pass of a chunked delta rule attention mechanism,
-optimized with Triton kernels for GPU acceleration. It includes functions for forward propagation,
-backward gradient computation, and integration with PyTorch's autograd system.
-
-该文件实现了分块 Delta Rule 注意力机制的前向与反向传播，
-使用 Triton 内核进行 GPU 加速优化。包括前向传播、梯度反向传播函数，
-并集成了 PyTorch 的自动求导系统。
-"""
-
 import warnings
 from typing import Optional
 
 import torch
 import triton
 
-# 导入内核实现模块 / Import kernel implementation modules
 from .torch_kernel.chunk_A_bwd import chunk_dplr_bwd_dqk_intra
 from .torch_kernel.chunk_A_fwd import chunk_dplr_fwd_intra
 from .torch_kernel.chunk_h_bwd import chunk_dplr_bwd_dhu
 from .torch_kernel.chunk_h_fwd import chunk_dplr_fwd_h
+
 from .torch_kernel.chunk_o_bwd import (
     chunk_dplr_bwd_dAu,
     chunk_dplr_bwd_dv,
@@ -37,11 +26,6 @@ from .get_torch_devices_info import (
 
 
 def cast(x, dtype):
-    """
-    Cast tensor x to specified dtype if not already in that format.
-
-    如果张量 x 不是目标数据类型，则将其转换为目标类型。
-    """
     if x is None or x.dtype == dtype:
         return x
     return x.to(dtype)
@@ -59,27 +43,6 @@ def chunk_dplr_fwd(
     output_final_state: bool = True,
     chunk_size: int = 16,
 ):
-    """
-    Forward pass of chunked delta rule attention.
-
-    分块 Delta Rule 注意力机制的前向传播。
-
-    Args:
-        q (torch.Tensor): Queries tensor [B, T, H, K]
-        k (torch.Tensor): Keys tensor [B, T, H, K]
-        v (torch.Tensor): Values tensor [B, T, H, V]
-        a (torch.Tensor): Activations tensor [B, T, H, K]
-        b (torch.Tensor): Betas tensor [B, T, H, K]
-        gk (torch.Tensor): Log decay tensor [B, T, H, K]
-        scale (float): Scale factor for attention scores
-        initial_state (Optional[torch.Tensor]): Initial state for recurrent processing
-        output_final_state (bool): Whether to return final state
-        chunk_size (int): Chunk size for processing
-
-    Returns:
-        o (torch.Tensor): Output tensor [B, T, H, V]
-        final_state (Optional[torch.Tensor]): Final state if requested
-    """
     T = q.shape[1]
     BT = min(chunk_size, max(triton.next_power_of_2(T), 16))
     gi, ge = chunk_rwkv6_fwd_cumsum(gk, BT)
@@ -97,11 +60,11 @@ def chunk_dplr_fwd(
 
     del ge
 
-    # Compute WY representation
+    # A_ab, A_ak, gi, ge torch.float32
+    # A_qk, A_qb, qg, kg, ag, bg, dtype=q.dtype, eg: bf16
     w, u, _ = prepare_wy_repr_fwd(ag=ag, A_ab=A_ab, A_ak=A_ak, v=v, chunk_size=BT)
 
     del A_ab, A_ak
-
     h, v_new, final_state = chunk_dplr_fwd_h(
         kg=kg,
         bg=bg,
@@ -137,33 +100,7 @@ def chunk_dplr_bwd(
     dht,
     BT: int = 16,
 ):
-    """
-    Backward pass of chunked delta rule attention.
-
-    分块 Delta Rule 注意力机制的反向传播。
-
-    Args:
-        q (torch.Tensor): Queries tensor [B, T, H, K]
-        k (torch.Tensor): Keys tensor [B, T, H, K]
-        v (torch.Tensor): Values tensor [B, T, H, V]
-        a (torch.Tensor): Activations tensor [B, T, H, K]
-        b (torch.Tensor): Betas tensor [B, T, H, K]
-        gk (torch.Tensor): Log decay tensor [B, T, H, K]
-        initial_state (torch.Tensor): Initial state for recurrent processing
-        scale (float): Scale factor for attention scores
-        do (torch.Tensor): Gradient of outputs
-        dht (torch.Tensor): Gradient of final hidden state
-        BT (int): Chunk size for processing
-
-    Returns:
-        dq (torch.Tensor): Gradient of queries
-        dk (torch.Tensor): Gradient of keys
-        dv (torch.Tensor): Gradient of values
-        da (torch.Tensor): Gradient of activations
-        db (torch.Tensor): Gradient of betas
-        dgk (torch.Tensor): Gradient of log decays
-        dh0 (torch.Tensor): Gradient of initial state
-    """
+    # ******* start recomputing everything, otherwise i believe the gpu memory will be exhausted *******
     gi, ge = chunk_rwkv6_fwd_cumsum(gk, BT)
     A_ab, A_qk, A_ak, A_qb, qg, kg, ag, bg = chunk_dplr_fwd_intra(
         q=q,
@@ -183,6 +120,9 @@ def chunk_dplr_bwd(
         kg=kg, bg=bg, v=v, w=w, u=u, gk=gi, initial_state=initial_state, chunk_size=BT
     )
     del u
+    # ******* end of recomputation *******
+    # A_ak, A_ab_inv, gi, ge torch.float32
+    # A_qk, A_qb, qg, kg, ag, bg, v_new dtype=q.dtype, eg: bf16
 
     dv_new_intra, dA_qk, dA_qb = chunk_dplr_bwd_dAu(
         v=v, v_new=v_new, do=do, A_qb=A_qb, scale=scale, chunk_size=BT
@@ -211,7 +151,7 @@ def chunk_dplr_bwd(
         do=do,
         h=h,
         dh=dh,
-        dv=dv,
+        dv=dv_new,
         w=w,
         gk=gi,
         chunk_size=BT,
@@ -282,11 +222,6 @@ class ChunkDPLRDeltaRuleFunction(torch.autograd.Function):
         output_final_state: bool = True,
         cu_seqlens: Optional[torch.LongTensor] = None,
     ):
-        """
-        Forward function with autograd support.
-
-        支持自动求导的前向函数。
-        """
         chunk_size = 16
         o, final_state = chunk_dplr_fwd(
             q=q,
@@ -310,11 +245,6 @@ class ChunkDPLRDeltaRuleFunction(torch.autograd.Function):
     @input_guard
     @autocast_custom_bwd
     def backward(ctx, do: torch.Tensor, dht: torch.Tensor):
-        """
-        Backward function with autograd support.
-
-        支持自动求导的反向函数。
-        """
         q, k, v, a, b, gk, initial_state = ctx.saved_tensors
         BT = ctx.chunk_size
         cu_seqlens = ctx.cu_seqlens
@@ -349,10 +279,6 @@ def chunk_dplr_delta_rule(
     cu_seqlens: Optional[torch.LongTensor] = None,
 ):
     r"""
-    Main interface function for chunked delta rule attention.
-
-    分块 Delta Rule 注意力机制的主要接口函数。
-
     Args:
         q (torch.Tensor):
             queries of shape `[B, T, H, K]`
@@ -435,10 +361,37 @@ def chunk_rwkv7(
     output_final_state: bool = True,
 ):
     """
-    Interface function for RWKV-7 attention.
-
-    RWKV-7 注意力机制的接口函数。
+    Args:
+        r (torch.Tensor):
+            r of shape `[B, H, T, K]` .
+        k (torch.Tensor):
+            k of shape `[B, H, T, K]` .
+        v (torch.Tensor):
+            v of shape `[B, H, T, V]` if `head_first=True` else `[B, T, H, V]`.
+        a (torch.Tensor):
+            a of shape `[B, H, T, K]` .
+        b (torch.Tensor):
+            b of shape `[B, H, T, K]` .
+        w (torch.Tensor):
+            decay of shape `[B, H, T, K]` , kernel
+            will apply log_w = -torch.exp(w)
+        log_w (torch.Tensor):
+            log decay of shape `[B, H, T, K]` .
+        scale (float):
+            scale of the attention.
+        initial_state (Optional[torch.Tensor]):
+            Initial state of shape `[N, H, K, V]` for `N` input sequences.
+            For equal-length input sequences, `N` equals the batch size `B`.
+            Default: `None`.
+        output_final_state (Optional[bool]):
+            Whether to output the final state of shape `[N, H, K, V]`. Default: `False`.
+        cu_seqlens (torch.LongTensor):
+            Cumulative sequence lengths of shape `[N+1]` used for variable-length training,
+            consistent with the FlashAttention API.
+        head_first (bool):
+            whether to use head first. Recommended to be False to avoid extra transposes.
     """
+
     if w is not None:
         log_w = -torch.exp(w)
     else:
@@ -458,11 +411,6 @@ def chunk_rwkv7(
 
 
 def transpose_head(x, head_first):
-    """
-    Transpose between head-first and time-first formats.
-
-    在 head-first 和 time-first 格式之间转置。
-    """
     if head_first:
         x = torch.permute(x, dims=(0, 2, 1, 3))
     out = cast(x, torch.bfloat16).contiguous()
@@ -480,11 +428,6 @@ def generalized_delta_rule(
     output_final_state: bool = True,
     head_first: bool = False,
 ):
-    """
-    Generalized delta rule attention interface.
-
-    泛化 Delta Rule 注意力机制接口。
-    """
     dtype = r.dtype
     r = transpose_head(r, head_first)
     k = transpose_head(k, head_first)
@@ -492,30 +435,16 @@ def generalized_delta_rule(
     a = transpose_head(a, head_first)
     b = transpose_head(b, head_first)
     w = transpose_head(w, head_first)
-    if w.device.type == "cuda":
-        out, state = chunk_rwkv7(
-            r=r,
-            k=k,
-            v=v,
-            a=a,
-            b=b,
-            w=w,
-            initial_state=initial_state,
-            output_final_state=output_final_state,
-        )
-    else:
-        from .native_keras_op import generalized_delta_rule
-
-        out, state = generalized_delta_rule(
-            r=r,
-            k=k,
-            v=v,
-            a=a,
-            b=b,
-            w=w,
-            initial_state=initial_state,
-            output_final_state=output_final_state,
-        )
+    out, state = chunk_rwkv7(
+        r=r,
+        k=k,
+        v=v,
+        a=a,
+        b=b,
+        w=w,
+        initial_state=initial_state,
+        output_final_state=output_final_state,
+    )
     out = transpose_head(out, head_first)
     if output_final_state:
         return out, cast(state, dtype)

{rwkv_ops-0.1.0.dist-info → rwkv_ops-0.1.1.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: rwkv-ops
-Version: 0.1.0
+Version: 0.1.1
 Home-page: https://github.com/your-org/rwkv_ops
 License: Apache 2.0
 Keywords: rwkv attention cuda triton pytorch jax
@@ -30,7 +30,8 @@ Requires-Dist: keras
 > 注意：本库依赖 `keras`。
 
 ---
-
+## 安装方法
+pip install rwkv_ops
 ## 环境变量
 
 | 变量名 | 含义 | 取值 | 默认值 | 优先级 |

{rwkv_ops-0.1.0.dist-info → rwkv_ops-0.1.1.dist-info}/RECORD

@@ -1,16 +1,16 @@
-rwkv_ops/__init__.py,sha256=Qp3EFYuuQB19ND6yvwxbczvGgH98hZcWr8_boIf_5I8,725
-rwkv_ops/rwkv7_kernel/__init__.py,sha256=nG_loc60Vp_ZQGLvwWFJYUIAsNCt-ldUYJnWdmVwyPc,5681
-rwkv_ops/rwkv7_kernel/get_jax_devices_info.py,sha256=fC_bISkSrKk1DvDNEaAdThmDZElz7T1BGWRl6EOXd6M,6005
-rwkv_ops/rwkv7_kernel/get_torch_devices_info.py,sha256=3Sgzw_-9VElBecmQ3TbW4y8GI9QZfwlP6BMb7CovlQ8,7380
-rwkv_ops/rwkv7_kernel/jax_op.py,sha256=xOg2YbPPHkXEJ59sOo1_stOrghJQPuj2nhXrqw4GSM4,9109
+rwkv_ops/__init__.py,sha256=zhiKsTn4RGCGGy_0VIZUgHjPZK9XlEJHy39bNkwPnH8,747
+rwkv_ops/rwkv7_kernel/__init__.py,sha256=k88BFK_NtUhG_27rK0_b48JCuEEXMb9_L9jGE50astc,6034
+rwkv_ops/rwkv7_kernel/get_jax_devices_info.py,sha256=cMIaNED7d1PvYNSyq8wNI3G7wNvcgdUj9HWRBLuSVM8,6004
+rwkv_ops/rwkv7_kernel/get_torch_devices_info.py,sha256=ZL_rAM6lHB4nTOOU28Xm08qptfuIoijOMi_xwJG3KCo,7380
+rwkv_ops/rwkv7_kernel/jax_op.py,sha256=tyMxvk_EblDaGsePpxw3AhELvolp7LeE5NopUhKw1R0,9107
 rwkv_ops/rwkv7_kernel/native_keras_op.py,sha256=QPrXLbqw0chipQg_0jepRp2U19BYpBBFdKZWyaDNNoc,2488
-rwkv_ops/rwkv7_kernel/torch_op.py,sha256=elbv2SOHs9fzsO0FGxQ6ug7HVsjTZAE2oIM_ZOOY4JM,14451
+rwkv_ops/rwkv7_kernel/torch_op.py,sha256=d6VQM7SS5ynQ_YTmqHzDIn2MLiXkYcMiSJD2eXEkTSg,12277
 rwkv_ops/rwkv7_kernel/jax_kernel/__init__.py,sha256=uHsf_1qrtRK62IvhLuzefHGPWpHXmw1p0tqmwlHcptk,346
 rwkv_ops/rwkv7_kernel/jax_kernel/chunk_A_bwd.py,sha256=2Voq1Bdzn0DFloiLvwINBk7akmxRWIqXIQeyafrJJGg,2138
 rwkv_ops/rwkv7_kernel/jax_kernel/chunk_A_fwd.py,sha256=rhmglqHIIww7yPzaSBEp9ISxhhxoUbMtV51AUDyhUd8,1425
 rwkv_ops/rwkv7_kernel/jax_kernel/chunk_h_bwd.py,sha256=JDfVZsMb8yMlMN3sKT3i3l3y1YQiQkyUjnSNyan5Fqc,1888
 rwkv_ops/rwkv7_kernel/jax_kernel/chunk_h_fwd.py,sha256=g8b_81rIIjxeknYiklRGnox24rAvEvfKRKT-5nI0Euo,1992
-rwkv_ops/rwkv7_kernel/jax_kernel/chunk_o_bwd.py,sha256=W5gutV5WXA1bscfuDFCYNoRiG526CjHy4eaD2VtYuh0,3357
+rwkv_ops/rwkv7_kernel/jax_kernel/chunk_o_bwd.py,sha256=gQnToi1e1GZCvjWsEdWx6WakUN4Lc0JfaBSsSXYdN84,3369
 rwkv_ops/rwkv7_kernel/jax_kernel/chunk_o_fwd.py,sha256=4SjQ_zTZvFxsBMeWOx0JGFg9EQ4vllvEx30EcvSZJzI,853
 rwkv_ops/rwkv7_kernel/jax_kernel/cumsum.py,sha256=NoOh2_hA_rdH5bmaNNMAdCgVPfWvQpf-Q8BqF926jrw,667
 rwkv_ops/rwkv7_kernel/jax_kernel/wy_fast_bwd.py,sha256=PAMtE6wCW2Hz39oiHLGqhxY77csQAMYdNP2najDO_Jg,1407
@@ -20,7 +20,7 @@ rwkv_ops/rwkv7_kernel/torch_kernel/chunk_A_bwd.py,sha256=CWtotXkVvHz4-rkuOqWh6zK
 rwkv_ops/rwkv7_kernel/torch_kernel/chunk_A_fwd.py,sha256=4RJbyUTO23OxwH1rGVxeBiBVZKNHpPL_tJ7MFoDCIts,1475
 rwkv_ops/rwkv7_kernel/torch_kernel/chunk_h_bwd.py,sha256=zo6l0ZZUhXFu8wEFD76I0zSqFT9IXFKUKtyeaSwk380,1795
 rwkv_ops/rwkv7_kernel/torch_kernel/chunk_h_fwd.py,sha256=0ucN1U0EDTDqcyTPLLcsAX6FLTf2E_3toOY9p81gWYE,1858
-rwkv_ops/rwkv7_kernel/torch_kernel/chunk_o_bwd.py,sha256=CR6rK729Jfdt3JgZMpZxmALyTIYT6a7bmvllCm0GIxI,3281
+rwkv_ops/rwkv7_kernel/torch_kernel/chunk_o_bwd.py,sha256=ioPrS0NYQhpFk1j8rAxqtbwpx1CwjJQnrJEBDqVy-As,3283
 rwkv_ops/rwkv7_kernel/torch_kernel/chunk_o_fwd.py,sha256=54yoa3NpV64H-koURt-hUWpFHhUjwXpGvXPp2_ETCnw,825
 rwkv_ops/rwkv7_kernel/torch_kernel/cumsum.py,sha256=hQkpyaa0eUyB4V3UVks7l1_dHwOrbump0FZILityBKw,611
 rwkv_ops/rwkv7_kernel/torch_kernel/wy_fast_bwd.py,sha256=gk6QdoT1oq5B8Hp8Ak-SGqHm8CEj3MErUeWcRsaaOQM,1470
@@ -36,8 +36,8 @@ rwkv_ops/rwkv7_kernel/triton_kernel/cumsum.py,sha256=pRp_z587PrnpgRVpi031IndyjVI
 rwkv_ops/rwkv7_kernel/triton_kernel/utils.py,sha256=TNGlkwGq4t-TOcdVBk_N_vHPLzMFTu_F0V-O1RprIO4,553
 rwkv_ops/rwkv7_kernel/triton_kernel/wy_fast_bwd.py,sha256=szaG11q_WmpyhXi6aVWwzizvflCh5wND8wGA_V8afzA,5479
 rwkv_ops/rwkv7_kernel/triton_kernel/wy_fast_fwd.py,sha256=jbb19DUTHENU2RIOv_T4m_W1eXMqdRqG0XevIkBOhI4,9438
-rwkv_ops-0.1.0.dist-info/LICENSE.txt,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
-rwkv_ops-0.1.0.dist-info/METADATA,sha256=KRIeem9Ulrk0bwVyelirLyWXcGVLoa1BXmwgjUPRikk,3572
-rwkv_ops-0.1.0.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
-rwkv_ops-0.1.0.dist-info/top_level.txt,sha256=cVqoKE-WR_e2gHL87-6O4K1kG6-yTJGB2huyr6FmD2I,9
-rwkv_ops-0.1.0.dist-info/RECORD,,
+rwkv_ops-0.1.1.dist-info/LICENSE.txt,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
+rwkv_ops-0.1.1.dist-info/METADATA,sha256=g2e5rhSz-SFLzyj76FbShCNtgdWAjgTV0ukw3WYR2fo,3608
+rwkv_ops-0.1.1.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
+rwkv_ops-0.1.1.dist-info/top_level.txt,sha256=cVqoKE-WR_e2gHL87-6O4K1kG6-yTJGB2huyr6FmD2I,9
+rwkv_ops-0.1.1.dist-info/RECORD,,