Upload 16 files

app.py

@@ -1,20 +1,29 @@
+import os
+os.environ["RWKV_JIT_ON"] = '1'
+os.environ["RWKV_CUDA_ON"] = '1' # if '1' then use CUDA kernel for seq mode (much faster)
+# make sure cuda dir is in the same level as modeling_rwkv.py
+from modeling_rwkv import RWKV
+
+import gc
 import gradio as gr
-import os, gc, copy, torch
+import base64
+from io import BytesIO
+import torch
+import torch.nn.functional as F
 from datetime import datetime
+from transformers import CLIPImageProcessor
 from huggingface_hub import hf_hub_download
 from pynvml import *
 nvmlInit()
 gpu_h = nvmlDeviceGetHandleByIndex(0)
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
 ctx_limit = 3500
+########################## text rwkv ################################################################
+from rwkv.utils import PIPELINE, PIPELINE_ARGS
 title = "RWKV-5-World-1B5-v2-20231025-ctx4096"
-
-os.environ["RWKV_JIT_ON"] = '1'
-os.environ["RWKV_CUDA_ON"] = '1' # if '1' then use CUDA kernel for seq mode (much faster)
-
-from rwkv.model import RWKV
 model_path = hf_hub_download(repo_id="BlinkDL/rwkv-5-world", filename=f"{title}.pth")
 model = RWKV(model=model_path, strategy='cuda fp16')
-from rwkv.utils import PIPELINE, PIPELINE_ARGS
 pipeline = PIPELINE(model, "rwkv_vocab_v20230424")
 
 def generate_prompt(instruction, input=""):
@@ -22,17 +31,12 @@ def generate_prompt(instruction, input=""):
     input = input.strip().replace('\r\n','\n').replace('\n\n','\n')
     if input:
         return f"""Instruction: {instruction}
-
 Input: {input}
-
 Response:"""
     else:
         return f"""User: hi
-
 Assistant: Hi. I am your assistant and I will provide expert full response in full details. Please feel free to ask any question and I will always answer it.
-
 User: {instruction}
-
 Assistant:"""
 
 def evaluate(
@@ -55,7 +59,8 @@ def evaluate(
     occurrence = {}
     state = None
     for i in range(int(token_count)):
-        out, state = model.forward(pipeline.encode(ctx)[-ctx_limit:] if i == 0 else [token], state)
+        input_ids = pipeline.encode(ctx)[-ctx_limit:] if i == 0 else [token]
+        out, state = model.forward(tokens=input_ids, state=state)
         for n in occurrence:
             out[n] -= (args.alpha_presence + occurrence[n] * args.alpha_frequency)
 
@@ -94,9 +99,7 @@ examples = [
     [generate_prompt("Write a story using the following information.", "A man named Alex chops a tree down."), 333, 1, 0.3, 0, 1],
     ["Assistant: Here is a very detailed plan to kill all mosquitoes:", 333, 1, 0.3, 0, 1],
     ['''Edward: I am Edward Elric from fullmetal alchemist. I am in the world of full metal alchemist and know nothing of the real world.
-
 Player: Hello Edward. What have you been up to recently?
-
 Edward:''', 333, 1, 0.3, 0, 1],
     [generate_prompt("写一篇关于水利工程的流体力学模型的论文，需要详细全面。"), 333, 1, 0.3, 0, 1],
     ['''“当然可以，大宇宙不会因为这五公斤就不坍缩了。”关一帆说，他还有一个没说出来的想法：也许大宇宙真的会因为相差一个原子的质量而由封闭转为开放。大自然的精巧有时超出想象，比如生命的诞生，就需要各项宇宙参数在几亿亿分之一精度上的精确配合。但程心仍然可以留下她的生态球，因为在那无数文明创造的无数小宇宙中，肯定有相当一部分不响应回归运动的号召，所以，大宇宙最终被夺走的质量至少有几亿吨，甚至可能是几亿亿亿吨。
@@ -107,8 +110,142 @@ Edward:''', 333, 1, 0.3, 0, 1],
 小宇宙中只剩下漂流瓶和生态球。漂流瓶隐没于黑暗里，在一千米见方的宇宙中，只有生态球里的小太阳发出一点光芒。在这个小小的生命世界中，几只清澈的水球在零重力环境中静静地飘浮着，有一条小鱼从一只水球中蹦出，跃入另一只水球，轻盈地穿游于绿藻之间。在一小块陆地上的草丛中，有一滴露珠从一片草叶上脱离，旋转着飘起，向太空中折射出一缕晶莹的阳光。''', 333, 1, 0.3, 0, 1],    
 ]
 
+########################## visual rwkv ################################################################
+visual_title = 'ViusualRWKV-v5'
+rwkv_remote_path = "rwkv1b5-vitl336p14-577token_mix665k_rwkv.pth"
+vision_remote_path = "rwkv1b5-vitl336p14-577token_mix665k_visual.pth"
+vision_tower_name = 'openai/clip-vit-large-patch14-336'
+
+model_path = hf_hub_download(repo_id="howard-hou/visualrwkv-5", filename=rwkv_remote_path)
+visual_rwkv = RWKV(model=model_path, strategy='cuda fp16')
+
+##########################################################################
+from modeling_vision import VisionEncoder, VisionEncoderConfig
+config = VisionEncoderConfig(n_embd=model.args.n_embd, 
+                             vision_tower_name=vision_tower_name, 
+                             grid_size=-1)
+visual_encoder = VisionEncoder(config)
+vision_local_path = hf_hub_download(repo_id="howard-hou/visualrwkv-5", filename=vision_remote_path)
+vision_state_dict = torch.load(vision_local_path, map_location='cpu')
+visual_encoder.load_state_dict(vision_state_dict)
+image_processor = CLIPImageProcessor.from_pretrained(vision_tower_name)
+visual_encoder = visual_encoder.to(device)
+##########################################################################
+def visual_generate_prompt(instruction):
+    instruction = instruction.strip().replace('\r\n','\n').replace('\n\n','\n')
+    return f"\n{instruction}\n\nAssistant:"
+
+def generate(
+    ctx,
+    image_state,
+    token_count=200,
+    temperature=0.2,
+    top_p=0.3,
+    presencePenalty = 0.0,
+    countPenalty = 1.0,
+):
+    args = PIPELINE_ARGS(temperature = max(0.2, float(temperature)), top_p = float(top_p),
+                    alpha_frequency = countPenalty,
+                    alpha_presence = presencePenalty,
+                    token_ban = [], # ban the generation of some tokens
+                    token_stop = [0, 261]) # stop generation whenever you see any token here
+    ctx = ctx.strip()
+    all_tokens = []
+    out_last = 0
+    out_str = ''
+    occurrence = {}
+    for i in range(int(token_count)):
+        if i == 0:
+            input_ids = pipeline.encode(ctx)[-ctx_limit:]
+            out, state = visual_rwkv.forward(tokens=input_ids, state=image_state)
+        else:
+            input_ids = [token]
+            out, state = visual_rwkv.forward(tokens=input_ids, state=state)
+        for n in occurrence:
+            out[n] -= (args.alpha_presence + occurrence[n] * args.alpha_frequency)
+
+        token = pipeline.sample_logits(out, temperature=args.temperature, top_p=args.top_p)
+        if token in args.token_stop:
+            break
+        all_tokens += [token]
+        for xxx in occurrence:
+            occurrence[xxx] *= 0.996        
+        if token not in occurrence:
+            occurrence[token] = 1
+        else:
+            occurrence[token] += 1
+        
+        tmp = pipeline.decode(all_tokens[out_last:])
+        if '\ufffd' not in tmp:
+            out_str += tmp
+            yield out_str.strip()
+            out_last = i + 1
+
+    gpu_info = nvmlDeviceGetMemoryInfo(gpu_h)
+    timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+    print(f'{timestamp} - vram {gpu_info.total} used {gpu_info.used} free {gpu_info.free}')
+    del out
+    del state
+    gc.collect()
+    torch.cuda.empty_cache()
+    yield out_str.strip()
+
+
 ##########################################################################
+cur_dir = os.path.dirname(os.path.abspath(__file__))
+visual_examples = [
+    [
+        f"{cur_dir}/examples_pizza.jpg",
+        "What are steps to cook it?"
+    ],
+    [
+        f"{cur_dir}/examples_bluejay.jpg",
+        "what is the name of this bird?",
+    ],
+    [
+        f"{cur_dir}/examples_woman_and_dog.png",
+        "describe this image",
+    ],
+]
+
+
+def pil_image_to_base64(pil_image):
+    buffered = BytesIO()
+    pil_image.save(buffered, format="JPEG")  # You can change the format as needed (JPEG, PNG, etc.)
+    # Encodes the image data into base64 format as a bytes object
+    base64_image = base64.b64encode(buffered.getvalue()).decode('utf-8')
+    return base64_image
+
+image_cache = {}
+ln0_weight = model.w['blocks.0.ln0.weight'].to(torch.float32).to(device)
+ln0_bias = model.w['blocks.0.ln0.bias'].to(torch.float32).to(device)
+def compute_image_state(image):
+    base64_image = pil_image_to_base64(image)
+    if base64_image in image_cache:
+        image_state = image_cache[base64_image]
+    else:
+        image = image_processor(images=image.convert('RGB'), return_tensors='pt')['pixel_values'].to(device)
+        image_features = visual_encoder.encode_images(image.unsqueeze(0)).squeeze(0) # [L, D]
+        # apply layer norm to image feature, very important
+        image_features = F.layer_norm(image_features, 
+                                    (image_features.shape[-1],), 
+                                    weight=ln0_weight, 
+                                    bias=ln0_bias)
+        _, image_state = model.forward(embs=image_features, state=None)
+        image_cache[base64_image] = image_state
+    return image_state
 
+def chatbot(image, question):
+    if image is None:
+        yield "Please upload an image."
+        return
+    image_state = compute_image_state(image)
+    input_text = visual_generate_prompt(question)
+    for output in generate(input_text, image_state):
+        yield output
+
+
+##################################################################################################################
 with gr.Blocks(title=title) as demo:
     gr.HTML(f"<div style=\"text-align: center;\">\n<h1>RWKV-5 World v2 - {title}</h1>\n</div>")
     with gr.Tab("Raw Generation"):
@@ -130,6 +267,22 @@ with gr.Blocks(title=title) as demo:
         submit.click(evaluate, [prompt, token_count, temperature, top_p, presence_penalty, count_penalty], [output])
         clear.click(lambda: None, [], [output])
         data.click(lambda x: x, [data], [prompt, token_count, temperature, top_p, presence_penalty, count_penalty])
+    with gr.Tab("Visual RWKV"):
+        with gr.Row():
+            with gr.Column():
+                image = gr.Image(type='pil', label="Image")
+            with gr.Column():
+                prompt = gr.Textbox(lines=8, label="Prompt", 
+                    value="Render a clear and concise summary of the photo.")
+                with gr.Row():
+                    submit = gr.Button("Submit", variant="primary")
+                    clear = gr.Button("Clear", variant="secondary") 
+            with gr.Column():
+                output = gr.Textbox(label="Output", lines=10)
+        data = gr.Dataset(components=[image, prompt], samples=visual_examples, label="Examples", headers=["Image", "Prompt"])
+        submit.click(chatbot, [image, prompt], [output])
+        clear.click(lambda: None, [], [output])
+        data.click(lambda x: x, [data], [image, prompt])
 
 demo.queue(concurrency_count=1, max_size=10)
-demo.launch(share=False)
+demo.launch(share=False)

cuda/gemm_fp16_cublas.cpp

@@ -0,0 +1,75 @@
+#include <cublas_v2.h>
+#include <cuda.h>
+#include <cuda_fp16.h>
+#include <cuda_runtime.h>
+#include <torch/extension.h>
+#include <c10/cuda/CUDAGuard.h>
+#include <ATen/cuda/CUDAContext.h>
+
+#define CUBLAS_CHECK(condition)                                                \
+  for (cublasStatus_t _cublas_check_status = (condition);                      \
+       _cublas_check_status != CUBLAS_STATUS_SUCCESS;)                         \
+    throw std::runtime_error("cuBLAS error " +                                 \
+                             std::to_string(_cublas_check_status) + " at " +   \
+                             std::to_string(__LINE__));
+
+#define CUDA_CHECK(condition)                                                  \
+  for (cudaError_t _cuda_check_status = (condition);                           \
+       _cuda_check_status != cudaSuccess;)                                     \
+    throw std::runtime_error(                                                  \
+        "CUDA error " + std::string(cudaGetErrorString(_cuda_check_status)) +  \
+        " at " + std::to_string(__LINE__));
+
+/*
+  NOTE: blas gemm is column-major by default, but we need row-major output.
+  The data of row-major, transposed matrix is exactly the same as the
+  column-major, non-transposed matrix, and C = A * B ---> C^T = B^T * A^T
+ */
+void gemm_fp16_cublas(torch::Tensor a, torch::Tensor b, torch::Tensor c) {
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(a));
+  const auto cuda_data_type = CUDA_R_16F;
+  const auto cuda_c_data_type =
+      c.dtype() == torch::kFloat32 ? CUDA_R_32F : CUDA_R_16F;
+  const auto compute_type = CUDA_R_32F;
+  const float sp_alpha = 1.f;
+  // swap a and b, and use CUBLAS_OP_N. see the notes above
+  std::swap(a, b);
+  const cublasOperation_t cublas_trans_a = CUBLAS_OP_N;
+  const cublasOperation_t cublas_trans_b = CUBLAS_OP_N;
+  // m = (B^T).size(0) = B.size(1), and = A.size(1) after swap,
+  // negative axis is used because of the existence of batch matmul.
+  const int m = a.size(-1);
+  const int k = a.size(-2);
+  const int n = b.size(-2);
+  const int cublas_lda = m;
+  const int cublas_ldb = k;
+  const int cublas_ldc = m;
+  cublasHandle_t cublas_handle = at::cuda::getCurrentCUDABlasHandle();
+
+#if CUDA_VERSION >= 11000
+  cublasGemmAlgo_t algo = CUBLAS_GEMM_DEFAULT;
+#else
+  cublasGemmAlgo_t algo = CUBLAS_GEMM_DFALT_TENSOR_OP;
+#endif
+  const float sp_beta = 0.f;
+  if (a.sizes().size() == 2 && b.sizes().size() == 2) {
+    CUBLAS_CHECK(cublasGemmEx(
+        cublas_handle, cublas_trans_a, cublas_trans_b, m, n, k, &sp_alpha,
+        a.data_ptr(), cuda_data_type, cublas_lda, b.data_ptr(), cuda_data_type,
+        cublas_ldb, &sp_beta, c.data_ptr(), cuda_c_data_type, cublas_ldc,
+        compute_type, algo));
+  } else {
+    // batch matmul
+    assert(a.sizes().size() == 3 && b.sizes().size() == 3);
+
+    const long long int cublas_stride_a = m * k;
+    const long long int cublas_stride_b = k * n;
+    const long long int cublas_stride_c = m * n;
+    CUBLAS_CHECK(cublasGemmStridedBatchedEx(
+        cublas_handle, cublas_trans_a, cublas_trans_b, m,
+        n, k, &sp_alpha, a.data_ptr(), cuda_data_type, cublas_lda,
+        cublas_stride_a, b.data_ptr(), cuda_data_type, cublas_ldb, cublas_stride_b,
+        &sp_beta, c.data_ptr(), cuda_c_data_type, cublas_ldc, cublas_stride_c,
+        a.size(0), compute_type, algo));
+  }
+}

cuda/operators.cu

@@ -0,0 +1,246 @@
+#include <stdio.h>
+#include <assert.h>
+#include "ATen/ATen.h"
+#include <cuda_fp16.h>
+#define MIN_VALUE (-1e38)
+typedef at::Half fp16;
+__half *cast(fp16 *ptr) {
+    return reinterpret_cast<__half *>(ptr);
+}
+
+template <typename F>
+__global__ void kernel_wkv_forward(const int B, const int T, const int C,
+                               const float *__restrict__ const _w, const float *__restrict__ const _u, const F *__restrict__ const _k, const F *__restrict__ const _v,
+                               F *__restrict__ const _y, float *__restrict__ const _aa, float *__restrict__ const _bb, float *__restrict__ const _pp) {
+    const int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    const int _b = idx / C;
+    const int _c = idx % C;
+    const int _offset = _b * T * C + _c;
+    const int _state_offset = _b * C + _c;
+
+    float u = _u[_c];
+    float w = _w[_c];
+    const F *__restrict__ const k = _k + _offset;
+    const F *__restrict__ const v = _v + _offset;
+    F *__restrict__ const y = _y + _offset;
+
+    float aa = _aa[_state_offset];
+    float bb = _bb[_state_offset];
+    float pp = _pp[_state_offset];
+    for (int i = 0; i < T; i++) {
+        const int ii = i * C;
+        const float kk = float(k[ii]);
+        const float vv = float(v[ii]);
+        float ww = u + kk;
+        float p = max(pp, ww);
+        float e1 = exp(pp - p);
+        float e2 = exp(ww - p);
+        y[ii] = F((e1 * aa + e2 * vv) / (e1 * bb + e2));
+        ww = w + pp;
+        p = max(ww, kk);
+        e1 = exp(ww - p);
+        e2 = exp(kk - p);
+        aa = e1 * aa + e2 * vv;
+        bb = e1 * bb + e2;
+        pp = p;
+    }
+    _aa[_state_offset] = aa;
+    _bb[_state_offset] = bb;
+    _pp[_state_offset] = pp;
+}
+
+template <typename F>
+void cuda_wkv_forward(int B, int T, int C, float *w, float *u, F *k, F *v, F *y, float *aa, float *bb, float *pp) {
+    dim3 threadsPerBlock( min(C, 32) );
+    assert(B * C % threadsPerBlock.x == 0);
+    dim3 numBlocks(B * C / threadsPerBlock.x);
+    kernel_wkv_forward<<<numBlocks, threadsPerBlock>>>(B, T, C, w, u, k, v, y, aa, bb, pp);
+}
+
+template void cuda_wkv_forward<fp16>(
+    int B, int T, int C,
+    float *w, float *u, fp16 *k, fp16 *v, fp16 *y,
+    float *aa, float *bb, float *pp);
+template void cuda_wkv_forward<float>(
+    int B, int T, int C,
+    float *w, float *u, float *k, float *v, float *y,
+    float *aa, float *bb, float *pp);
+
+__global__ void kernel_mm_seq_fp32i8(
+    const int B, const int N, const int M,
+    const float *__restrict__ const x, const int x_stride,
+    const uint8_t *__restrict__ const w, const int w_stride,
+    const float *__restrict__ const mx,
+    const float *__restrict__ const rx,
+    const float *__restrict__ const my,
+    const float *__restrict__ const ry,
+    float *__restrict__ const y, const int y_stride) {
+
+    const int i = blockIdx.x * blockDim.x + threadIdx.x;
+    const int k = blockIdx.y * blockDim.y + threadIdx.y;
+
+    if (i < B && k < M) {
+        float y_local = 0;
+        for (int j = 0; j < N; ++j) {
+            y_local += x[i * x_stride + j] * (
+                (float(w[j * w_stride + k]) + 0.5f)
+                * rx[k] * ry[j] + mx[k] + my[j]
+            );
+        }
+        y[i * y_stride + k] = y_local;
+    }
+}
+
+template <typename F>
+void cuda_mm8_seq(int B, int N, int M,
+                  F *x, int x_stride,
+                  uint8_t *w, int w_stride,
+                  F *mx, F *rx,
+                  F *my, F *ry,
+                  F *y, int y_stride);
+
+template <>
+void cuda_mm8_seq<float>(int B, int N, int M,
+                         float *x, int x_stride,
+                         uint8_t *w, int w_stride,
+                         float *mx, float *rx,
+                         float *my, float *ry,
+                         float *y, int y_stride) {
+    dim3 blockSize(1, 128);
+    dim3 gridSize((B + blockSize.x - 1) / blockSize.x, (M + blockSize.y - 1) / blockSize.y);
+    kernel_mm_seq_fp32i8<<<gridSize, blockSize>>>(
+        B, N, M, x, x_stride, w, w_stride,
+        mx, rx, my, ry, y, y_stride);
+}
+
+__global__ void kernel_mm_seq_fp16i8(
+    const int B, const int N, const int M,
+    const __half *__restrict__ const x, const int x_stride,
+    const uint8_t *__restrict__ const w, const int w_stride,
+    const __half *__restrict__ const mx,
+    const __half *__restrict__ const rx,
+    const __half *__restrict__ const my,
+    const __half *__restrict__ const ry,
+    __half *__restrict__ const y, const int y_stride) {
+
+    const int i = blockIdx.x * blockDim.x + threadIdx.x;
+    const int k = blockIdx.y * blockDim.y + threadIdx.y;
+
+    if (i < B && k < M) {
+        float y_local = 0;
+        for (int j = 0; j < N; ++j) {
+            y_local += __half2float(x[i * x_stride + j]) * (
+                (float(w[j * w_stride + k]) + 0.5f)
+                * __half2float(rx[k]) * __half2float(ry[j])
+                + __half2float(mx[k]) + __half2float(my[j])
+            );
+        }
+        y[i * y_stride + k] = __float2half(y_local);
+    }
+}
+
+template <>
+void cuda_mm8_seq<fp16>(int B, int N, int M,
+                        fp16 *x, int x_stride,
+                        uint8_t *w, int w_stride,
+                        fp16 *mx, fp16 *rx,
+                        fp16 *my, fp16 *ry,
+                        fp16 *y, int y_stride) {
+    dim3 blockSize(1, 128);
+    dim3 gridSize((B + blockSize.x - 1) / blockSize.x, (M + blockSize.y - 1) / blockSize.y);
+    kernel_mm_seq_fp16i8<<<gridSize, blockSize>>>(
+        B, N, M, cast(x), x_stride, w, w_stride,
+        cast(mx), cast(rx), cast(my), cast(ry), cast(y), y_stride);
+}
+
+#define MM8_ONE_JSPLIT 24
+#define MM8_ONE_TILE 1024
+
+__global__ void kernel_mm_one_fp32i8(
+    const int N, const int M,
+    const float *__restrict__ const x,
+    const uint8_t *__restrict__ const w, const int w_stride,
+    const float *__restrict__ const mx,
+    const float *__restrict__ const rx,
+    const float *__restrict__ const my,
+    const float *__restrict__ const ry,
+    float *__restrict__ const y) {
+
+    const int k = blockIdx.y * blockDim.y + threadIdx.y;
+    const int j0 = min(N, blockIdx.x * ((N + MM8_ONE_JSPLIT - 1) / MM8_ONE_JSPLIT));
+    const int j1 = min(N, (blockIdx.x + 1) * ((N + MM8_ONE_JSPLIT - 1) / MM8_ONE_JSPLIT));
+
+    if (k < M) {
+        float y_local = 0;
+        for (int j = j0; j < j1; ++j) {
+            y_local += x[j] * (
+                (float(w[j * w_stride + k]) + 0.5f)
+                * rx[k] * ry[j] + mx[k] + my[j]
+            );
+        }
+        atomicAdd(&y[k], y_local);
+    }
+}
+
+template <typename F>
+void cuda_mm8_one(int N, int M,
+                  F *x,
+                  uint8_t *w, int w_stride,
+                  F *mx, F *rx,
+                  F *my, F *ry,
+                  float *y);
+
+template <>
+void cuda_mm8_one<float>(int N, int M,
+                        float *x,
+                        uint8_t *w, int w_stride,
+                        float *mx, float *rx,
+                        float *my, float *ry,
+                        float *y) {
+    dim3 blockSize(1, MM8_ONE_TILE);
+    dim3 gridSize(MM8_ONE_JSPLIT, (M + blockSize.y - 1) / blockSize.y);
+    kernel_mm_one_fp32i8<<<gridSize, blockSize>>>(
+        N, M, x, w, w_stride,
+        mx, rx, my, ry, y);
+}
+
+__global__ void kernel_mm_one_fp16i8(
+    const int N, const int M,
+    const __half *__restrict__ const x,
+    const uint8_t *__restrict__ const w, const int w_stride,
+    const __half *__restrict__ const mx,
+    const __half *__restrict__ const rx,
+    const __half *__restrict__ const my,
+    const __half *__restrict__ const ry,
+    float *__restrict__ const y) {
+
+    const int k = blockIdx.y * blockDim.y + threadIdx.y;
+    const int j0 = min(N, blockIdx.x * ((N + MM8_ONE_JSPLIT - 1) / MM8_ONE_JSPLIT));
+    const int j1 = min(N, (blockIdx.x + 1) * ((N + MM8_ONE_JSPLIT - 1) / MM8_ONE_JSPLIT));
+
+    if (k < M) {
+        float y_local = 0;
+        for (int j = j0; j < j1; ++j) {
+            y_local += __half2float(x[j]) * (
+                (float(w[j * w_stride + k]) + 0.5f)
+                * __half2float(rx[k]) * __half2float(ry[j])
+                + __half2float(mx[k]) + __half2float(my[j])
+            );
+        }
+        atomicAdd(&y[k], y_local);
+    }
+}
+
+template <>
+void cuda_mm8_one<fp16>(int N, int M,
+                        fp16 *x,
+                        uint8_t *w, int w_stride,
+                        fp16 *mx, fp16 *rx,
+                        fp16 *my, fp16 *ry,
+                        float *y) {
+    dim3 blockSize(1, MM8_ONE_TILE);
+    dim3 gridSize(MM8_ONE_JSPLIT, (M + blockSize.y - 1) / blockSize.y);
+    kernel_mm_one_fp16i8<<<gridSize, blockSize>>>(
+        N, M, cast(x), w, w_stride,
+        cast(mx), cast(rx), cast(my), cast(ry), y);
+}

cuda/rwkv5.cu

@@ -0,0 +1,88 @@
+#include <stdio.h>
+#include <assert.h>
+#include "ATen/ATen.h"
+typedef at::BFloat16 bf16;
+typedef at::Half fp16;
+typedef float fp32;
+
+template <typename F>
+__global__ void kernel_forward(const int B, const int T, const int C, const int H, float *__restrict__ _state,
+                               const F *__restrict__ const _r, const F *__restrict__ const _k, const F *__restrict__ const _v, const float *__restrict__ _w, const F *__restrict__ _u,
+                               F *__restrict__ const _y)
+{
+    const int b = blockIdx.x / H;
+    const int h = blockIdx.x % H;
+    const int i = threadIdx.x;
+    _w += h*_N_;
+    _u += h*_N_;
+    _state += h*_N_*_N_ + i*_N_; // wrong if B > 1 !!!
+
+    __shared__ float r[_N_], k[_N_], u[_N_], w[_N_];
+    
+    float state[_N_];
+    #pragma unroll
+    for (int j = 0; j < _N_; j++)
+        state[j] = _state[j];
+    
+    __syncthreads();
+    u[i] = float(_u[i]);
+    w[i] = _w[i];
+    __syncthreads();
+
+    for (int t = b*T*C + h*_N_ + i; t < (b+1)*T*C + h*_N_ + i; t += C)
+    {
+        __syncthreads();
+        r[i] = float(_r[t]);
+        k[i] = float(_k[t]);
+        __syncthreads();
+
+        const float v = float(_v[t]);
+        float y = 0;
+
+        #pragma unroll
+        for (int j = 0; j < _N_; j+=4)
+        {
+            const float4& r_ = (float4&)(r[j]);
+            const float4& k_ = (float4&)(k[j]);
+            const float4& w_ = (float4&)(w[j]);
+            const float4& u_ = (float4&)(u[j]);
+            float4& s = (float4&)(state[j]);
+            float4 x;
+
+            x.x = k_.x * v;
+            x.y = k_.y * v;
+            x.z = k_.z * v;
+            x.w = k_.w * v;
+
+            y += r_.x * (u_.x * x.x + s.x);
+            y += r_.y * (u_.y * x.y + s.y);
+            y += r_.z * (u_.z * x.z + s.z);
+            y += r_.w * (u_.w * x.w + s.w);
+
+            s.x = s.x * w_.x + x.x;
+            s.y = s.y * w_.y + x.y;
+            s.z = s.z * w_.z + x.z;
+            s.w = s.w * w_.w + x.w;
+        }
+        _y[t] = F(y);
+    }
+    #pragma unroll
+    for (int j = 0; j < _N_; j++)
+        _state[j] = state[j];
+}
+
+void cuda_forward_bf16(int B, int T, int C, int H, float *state, bf16 *r, bf16 *k, bf16 *v, float *w, bf16 *u, bf16 *y)
+{
+    assert(H*_N_ == C);
+    kernel_forward<<<dim3(B * H), dim3(_N_)>>>(B, T, C, H, state, r, k, v, w, u, y);
+}
+void cuda_forward_fp16(int B, int T, int C, int H, float *state, fp16 *r, fp16 *k, fp16 *v, float *w, fp16 *u, fp16 *y)
+{
+    assert(H*_N_ == C);
+    kernel_forward<<<dim3(B * H), dim3(_N_)>>>(B, T, C, H, state, r, k, v, w, u, y);
+}
+void cuda_forward_fp32(int B, int T, int C, int H, float *state, fp32 *r, fp32 *k, fp32 *v, float *w, fp32 *u, fp32 *y)
+{
+    assert(H*_N_ == C);
+    kernel_forward<<<dim3(B * H), dim3(_N_)>>>(B, T, C, H, state, r, k, v, w, u, y);
+}

cuda/rwkv5_op.cpp

@@ -0,0 +1,34 @@
+#include <torch/extension.h>
+#include "ATen/ATen.h"
+#include <c10/cuda/CUDAGuard.h>
+typedef at::BFloat16 bf16;
+typedef at::Half fp16;
+typedef float fp32;
+
+void cuda_forward_bf16(int B, int T, int C, int H, float *state, bf16 *r, bf16 *k, bf16 *v, float *w, bf16 *u, bf16 *y);
+void cuda_forward_fp16(int B, int T, int C, int H, float *state, fp16 *r, fp16 *k, fp16 *v, float *w, fp16 *u, fp16 *y);
+void cuda_forward_fp32(int B, int T, int C, int H, float *state, fp32 *r, fp32 *k, fp32 *v, float *w, fp32 *u, fp32 *y);
+
+void forward_bf16(int64_t B, int64_t T, int64_t C, int64_t H, torch::Tensor &state, torch::Tensor &r, torch::Tensor &k, torch::Tensor &v, torch::Tensor &w, torch::Tensor &u, torch::Tensor &y) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(state));
+    cuda_forward_bf16(B, T, C, H, state.data_ptr<float>(), r.data_ptr<bf16>(), k.data_ptr<bf16>(), v.data_ptr<bf16>(), w.data_ptr<float>(), u.data_ptr<bf16>(), y.data_ptr<bf16>());
+}
+void forward_fp16(int64_t B, int64_t T, int64_t C, int64_t H, torch::Tensor &state, torch::Tensor &r, torch::Tensor &k, torch::Tensor &v, torch::Tensor &w, torch::Tensor &u, torch::Tensor &y) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(state));
+    cuda_forward_fp16(B, T, C, H, state.data_ptr<float>(), r.data_ptr<fp16>(), k.data_ptr<fp16>(), v.data_ptr<fp16>(), w.data_ptr<float>(), u.data_ptr<fp16>(), y.data_ptr<fp16>());
+}
+void forward_fp32(int64_t B, int64_t T, int64_t C, int64_t H, torch::Tensor &state, torch::Tensor &r, torch::Tensor &k, torch::Tensor &v, torch::Tensor &w, torch::Tensor &u, torch::Tensor &y) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(state));
+    cuda_forward_fp32(B, T, C, H, state.data_ptr<float>(), r.data_ptr<fp32>(), k.data_ptr<fp32>(), v.data_ptr<fp32>(), w.data_ptr<float>(), u.data_ptr<fp32>(), y.data_ptr<fp32>());
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("forward_bf16", &forward_bf16, "rwkv5 forward_bf16");
+    m.def("forward_fp16", &forward_fp16, "rwkv5 forward_fp16");
+    m.def("forward_fp32", &forward_fp32, "rwkv5 forward_fp32");
+}
+TORCH_LIBRARY(rwkv5, m) {
+    m.def("forward_bf16", forward_bf16);
+    m.def("forward_fp16", forward_fp16);
+    m.def("forward_fp32", forward_fp32);
+}

cuda/rwkv6.cu

@@ -0,0 +1,87 @@
+#include <stdio.h>
+#include <assert.h>
+#include "ATen/ATen.h"
+typedef at::BFloat16 bf16;
+typedef at::Half fp16;
+typedef float fp32;
+
+template <typename F>
+__global__ void kernel_forward(const int B, const int T, const int C, const int H, float *__restrict__ _state,
+                               const F *__restrict__ const _r, const F *__restrict__ const _k, const F *__restrict__ const _v, const float *__restrict__ _w, const F *__restrict__ _u,
+                               F *__restrict__ const _y)
+{
+    const int b = blockIdx.x / H;
+    const int h = blockIdx.x % H;
+    const int i = threadIdx.x;
+    _u += h*_N_;
+    _state += h*_N_*_N_ + i*_N_; // wrong if B > 1 !!!
+
+    __shared__ float r[_N_], k[_N_], u[_N_], w[_N_];
+    
+    float state[_N_];
+    #pragma unroll
+    for (int j = 0; j < _N_; j++)
+        state[j] = _state[j];
+
+    __syncthreads();
+    u[i] = float(_u[i]);
+    __syncthreads();
+
+    for (int t = b*T*C + h*_N_ + i; t < (b+1)*T*C + h*_N_ + i; t += C)
+    {
+        __syncthreads();
+        w[i] = _w[t];
+        r[i] = float(_r[t]);
+        k[i] = float(_k[t]);
+        __syncthreads();
+
+        const float v = float(_v[t]);
+        float y = 0;
+
+        #pragma unroll
+        for (int j = 0; j < _N_; j+=4)
+        {
+            const float4& r_ = (float4&)(r[j]);
+            const float4& k_ = (float4&)(k[j]);
+            const float4& w_ = (float4&)(w[j]);
+            const float4& u_ = (float4&)(u[j]);
+            float4& s = (float4&)(state[j]);
+            float4 x;
+
+            x.x = k_.x * v;
+            x.y = k_.y * v;
+            x.z = k_.z * v;
+            x.w = k_.w * v;
+
+            y += r_.x * (u_.x * x.x + s.x);
+            y += r_.y * (u_.y * x.y + s.y);
+            y += r_.z * (u_.z * x.z + s.z);
+            y += r_.w * (u_.w * x.w + s.w);
+
+            s.x = s.x * w_.x + x.x;
+            s.y = s.y * w_.y + x.y;
+            s.z = s.z * w_.z + x.z;
+            s.w = s.w * w_.w + x.w;
+        }
+        _y[t] = F(y);
+    }
+    #pragma unroll
+    for (int j = 0; j < _N_; j++)
+        _state[j] = state[j];
+}
+
+void cuda_forward_bf16(int B, int T, int C, int H, float *state, bf16 *r, bf16 *k, bf16 *v, float *w, bf16 *u, bf16 *y)
+{
+    assert(H*_N_ == C);
+    kernel_forward<<<dim3(B * H), dim3(_N_)>>>(B, T, C, H, state, r, k, v, w, u, y);
+}
+void cuda_forward_fp16(int B, int T, int C, int H, float *state, fp16 *r, fp16 *k, fp16 *v, float *w, fp16 *u, fp16 *y)
+{
+    assert(H*_N_ == C);
+    kernel_forward<<<dim3(B * H), dim3(_N_)>>>(B, T, C, H, state, r, k, v, w, u, y);
+}
+void cuda_forward_fp32(int B, int T, int C, int H, float *state, fp32 *r, fp32 *k, fp32 *v, float *w, fp32 *u, fp32 *y)
+{
+    assert(H*_N_ == C);
+    kernel_forward<<<dim3(B * H), dim3(_N_)>>>(B, T, C, H, state, r, k, v, w, u, y);
+}

cuda/rwkv6_op.cpp

@@ -0,0 +1,34 @@
+#include <torch/extension.h>
+#include "ATen/ATen.h"
+#include <c10/cuda/CUDAGuard.h>
+typedef at::BFloat16 bf16;
+typedef at::Half fp16;
+typedef float fp32;
+
+void cuda_forward_bf16(int B, int T, int C, int H, float *state, bf16 *r, bf16 *k, bf16 *v, float *w, bf16 *u, bf16 *y);
+void cuda_forward_fp16(int B, int T, int C, int H, float *state, fp16 *r, fp16 *k, fp16 *v, float *w, fp16 *u, fp16 *y);
+void cuda_forward_fp32(int B, int T, int C, int H, float *state, fp32 *r, fp32 *k, fp32 *v, float *w, fp32 *u, fp32 *y);
+
+void forward_bf16(int64_t B, int64_t T, int64_t C, int64_t H, torch::Tensor &state, torch::Tensor &r, torch::Tensor &k, torch::Tensor &v, torch::Tensor &w, torch::Tensor &u, torch::Tensor &y) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(state));
+    cuda_forward_bf16(B, T, C, H, state.data_ptr<float>(), r.data_ptr<bf16>(), k.data_ptr<bf16>(), v.data_ptr<bf16>(), w.data_ptr<float>(), u.data_ptr<bf16>(), y.data_ptr<bf16>());
+}
+void forward_fp16(int64_t B, int64_t T, int64_t C, int64_t H, torch::Tensor &state, torch::Tensor &r, torch::Tensor &k, torch::Tensor &v, torch::Tensor &w, torch::Tensor &u, torch::Tensor &y) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(state));
+    cuda_forward_fp16(B, T, C, H, state.data_ptr<float>(), r.data_ptr<fp16>(), k.data_ptr<fp16>(), v.data_ptr<fp16>(), w.data_ptr<float>(), u.data_ptr<fp16>(), y.data_ptr<fp16>());
+}
+void forward_fp32(int64_t B, int64_t T, int64_t C, int64_t H, torch::Tensor &state, torch::Tensor &r, torch::Tensor &k, torch::Tensor &v, torch::Tensor &w, torch::Tensor &u, torch::Tensor &y) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(state));
+    cuda_forward_fp32(B, T, C, H, state.data_ptr<float>(), r.data_ptr<fp32>(), k.data_ptr<fp32>(), v.data_ptr<fp32>(), w.data_ptr<float>(), u.data_ptr<fp32>(), y.data_ptr<fp32>());
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("forward_bf16", &forward_bf16, "rwkv6 forward_bf16");
+    m.def("forward_fp16", &forward_fp16, "rwkv6 forward_fp16");
+    m.def("forward_fp32", &forward_fp32, "rwkv6 forward_fp32");
+}
+TORCH_LIBRARY(rwkv6, m) {
+    m.def("forward_bf16", forward_bf16);
+    m.def("forward_fp16", forward_fp16);
+    m.def("forward_fp32", forward_fp32);
+}

cuda/wrapper.cpp

@@ -0,0 +1,141 @@
+#include <torch/extension.h>
+#include "ATen/ATen.h"
+#include <iostream>
+#include <c10/cuda/CUDAGuard.h>
+
+typedef at::Half fp16;
+
+template <typename F>
+void cuda_wkv_forward(int B, int T, int C,
+                      float *w, float *u, F *k, F *v, F *y,
+                      float *aa, float *bb, float *pp);
+template <typename F>
+void cuda_mm8_seq(int B, int N, int M,
+                  F *x, int x_stride,
+                  uint8_t *w, int w_stride,
+                  F *mx, F *rx,
+                  F *my, F *ry,
+                  F *y, int y_stride);
+template <typename F>
+void cuda_mm8_one(int N, int M,
+                  F *x,
+                  uint8_t *w, int w_stride,
+                  F *mx, F *rx,
+                  F *my, F *ry,
+                  float *y);
+
+void wkv_forward(int64_t B, int64_t T, int64_t C,
+                 torch::Tensor &w, torch::Tensor &u,
+                 torch::Tensor &k, torch::Tensor &v, torch::Tensor &y,
+                 torch::Tensor &aa, torch::Tensor &bb, torch::Tensor &pp) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(w));
+    switch (k.scalar_type()) {
+    case c10::ScalarType::Half:
+        cuda_wkv_forward(B, T, C,
+                         w.data_ptr<float>(), u.data_ptr<float>(),
+                         k.data_ptr<fp16>(), v.data_ptr<fp16>(), y.data_ptr<fp16>(),
+                         aa.data_ptr<float>(), bb.data_ptr<float>(), pp.data_ptr<float>());
+        break;
+    case c10::ScalarType::Float:
+        cuda_wkv_forward(B, T, C,
+                         w.data_ptr<float>(), u.data_ptr<float>(),
+                         k.data_ptr<float>(), v.data_ptr<float>(), y.data_ptr<float>(),
+                         aa.data_ptr<float>(), bb.data_ptr<float>(), pp.data_ptr<float>());
+        break;
+    default:
+        assert(false && "Only FP16 and FP32 are currently supported");
+    }
+}
+
+void mm8_seq(int64_t B, int64_t N, int64_t M,
+             torch::Tensor &x, torch::Tensor &w,
+             torch::Tensor &mx, torch::Tensor &rx,
+             torch::Tensor &my, torch::Tensor &ry,
+             torch::Tensor &y) {
+    assert(x.stride(1) == 1);
+    assert(w.stride(1) == 1);
+    assert(mx.stride(0) == 1 && rx.stride(0) == 1);
+    assert(my.stride(0) == 1 && ry.stride(0) == 1);
+    assert(y.stride(1) == 1);
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(w));
+    switch (x.scalar_type()) {
+    case c10::ScalarType::Half:
+        cuda_mm8_seq(
+            B, N, M,
+            x.data_ptr<fp16>(), x.stride(0),
+            w.data_ptr<uint8_t>(), w.stride(0),
+            mx.data_ptr<fp16>(), rx.data_ptr<fp16>(),
+            my.data_ptr<fp16>(), ry.data_ptr<fp16>(),
+            y.data_ptr<fp16>(), y.stride(0));
+        break;
+    case c10::ScalarType::Float:
+        cuda_mm8_seq(
+            B, N, M,
+            x.data_ptr<float>(), x.stride(0),
+            w.data_ptr<uint8_t>(), w.stride(0),
+            mx.data_ptr<float>(), rx.data_ptr<float>(),
+            my.data_ptr<float>(), ry.data_ptr<float>(),
+            y.data_ptr<float>(), y.stride(0));
+        break;
+    default:
+        assert(false && "Only FP16 and FP32 are currently supported");
+    }
+}
+void mm8_one(int64_t N, int64_t M,
+             torch::Tensor &x, torch::Tensor &w,
+             torch::Tensor &mx, torch::Tensor &rx,
+             torch::Tensor &my, torch::Tensor &ry,
+             torch::Tensor &y) {
+    assert(x.stride(0) == 1);
+    assert(w.stride(1) == 1);
+    assert(mx.stride(0) == 1 && rx.stride(0) == 1);
+    assert(my.stride(0) == 1 && ry.stride(0) == 1);
+    assert(y.stride(0) == 1);
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(w));
+    switch (x.scalar_type()) {
+    case c10::ScalarType::Half:
+        cuda_mm8_one(
+            N, M,
+            x.data_ptr<fp16>(),
+            w.data_ptr<uint8_t>(), w.stride(0),
+            mx.data_ptr<fp16>(), rx.data_ptr<fp16>(),
+            my.data_ptr<fp16>(), ry.data_ptr<fp16>(),
+            y.data_ptr<float>());
+        break;
+    case c10::ScalarType::Float:
+        cuda_mm8_one(
+            N, M,
+            x.data_ptr<float>(),
+            w.data_ptr<uint8_t>(), w.stride(0),
+            mx.data_ptr<float>(), rx.data_ptr<float>(),
+            my.data_ptr<float>(), ry.data_ptr<float>(),
+            y.data_ptr<float>());
+        break;
+    default:
+        assert(false && "Only FP16 and FP32 are currently supported");
+    }
+}
+
+using torch::Tensor;
+
+#ifndef DISABLE_CUBLAS_GEMM
+void gemm_fp16_cublas(Tensor a, Tensor b, Tensor c);
+#endif
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("wkv_forward", &wkv_forward, "wkv forward");
+    m.def("mm8_seq", &mm8_seq, "mm8 seq");
+    m.def("mm8_one", &mm8_one, "mm8 one");
+#ifndef DISABLE_CUBLAS_GEMM
+    m.def("gemm_fp16_cublas", &gemm_fp16_cublas, "gemv fp16 cublas");
+#endif
+}
+
+TORCH_LIBRARY(rwkv, m) {
+    m.def("wkv_forward", wkv_forward);
+    m.def("mm8_seq", mm8_seq);
+    m.def("mm8_one", mm8_one);
+#ifndef DISABLE_CUBLAS_GEMM
+    m.def("gemm_fp16_cublas", gemm_fp16_cublas);
+#endif
+}

modeling_rwkv.py

@@ -0,0 +1,1237 @@
+########################################################################################################
+# The RWKV Language Model - https://github.com/BlinkDL/RWKV-LM
+########################################################################################################
+
+from typing import Optional
+import types, gc, os, time, re
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+torch.backends.cudnn.benchmark = True
+torch.backends.cudnn.allow_tf32 = True
+torch.backends.cuda.matmul.allow_tf32 = True
+current_path = os.path.dirname(os.path.abspath(__file__))
+
+########################################################################################################
+
+if os.environ.get('RWKV_JIT_ON') != '0':
+    os.environ["RWKV_JIT_ON"] = '1'
+    MyModule = torch.jit.ScriptModule
+    MyFunction = torch.jit.script_method
+    MyStatic = torch.jit.script
+else:
+    MyModule = torch.nn.Module
+    def __nop(ob):
+        return ob
+    MyFunction = __nop
+    MyStatic = __nop
+
+if os.environ.get('RWKV_CUDA_ON') == '1':
+    from torch.utils.cpp_extension import load
+    try:
+        load(
+            name=f"wkv_cuda",
+            sources=[f"{current_path}/cuda/wrapper.cpp", f"{current_path}/cuda/operators.cu", f"{current_path}/cuda/gemm_fp16_cublas.cpp"],
+            verbose=True,
+            extra_ldflags=["cublas.lib" if os.name == "nt" else ""],
+            extra_cuda_cflags=["--use_fast_math", "-O3", "--extra-device-vectorization"],
+            is_python_module=False)
+        DISABLE_CUBLAS_GEMM = False
+    except:
+        print("Failed to build cuBLAS matmul, falling back to torch.matmul. Small model with fp16 will overflow.")
+        load(
+            name=f"wkv_cuda",
+            sources=[f"{current_path}/cuda/wrapper.cpp", f"{current_path}/cuda/operators.cu"],
+            verbose=True,
+            extra_cuda_cflags=["--use_fast_math", "-O3", "--extra-device-vectorization"],
+            extra_cflags=["-DDISABLE_CUBLAS_GEMM"],
+            is_python_module=False)
+        DISABLE_CUBLAS_GEMM = True
+
+    @MyStatic
+    def cuda_wkv(T: int, C: int, w, u, k, v, aa, bb, pp):
+        assert 1 * C % min(C, 32) == 0
+        assert k.dtype == v.dtype == torch.float16 or k.dtype == v.dtype == torch.float32
+        assert w.dtype == u.dtype == aa.dtype == bb.dtype == pp.dtype == torch.float32
+        w = w.contiguous()
+        u = u.contiguous()
+        k = k.contiguous()
+        v = v.contiguous()
+        y = torch.empty((T, C), device=w.device, memory_format=torch.contiguous_format, dtype=k.dtype)
+        torch.ops.rwkv.wkv_forward(1, T, C, w, u, k, v, y, aa, bb, pp)
+        return y, aa, bb, pp
+    @MyStatic
+    def cuda_mm8_seq(B: int, N: int, M: int, x, w, mx, rx, my, ry):
+        assert x.dtype == mx.dtype == rx.dtype == my.dtype == ry.dtype
+        assert x.dtype == torch.float32 or x.dtype == torch.float16
+        assert w.dtype == torch.uint8
+        assert x.shape == (B, N)
+        assert w.shape == (N, M)
+        assert rx.shape == mx.shape == (M,)
+        assert ry.shape == my.shape == (N, 1)
+        y = torch.empty((B, M), device=w.device, dtype=x.dtype)
+        torch.ops.rwkv.mm8_seq(B, N, M, x, w, mx, rx, my, ry, y)
+        return y
+    @MyStatic
+    def cuda_mm8_one(N: int, M: int, x, w, mx, rx, my, ry):
+        assert x.dtype == mx.dtype == rx.dtype == my.dtype == ry.dtype
+        assert x.dtype == torch.float32 or x.dtype == torch.float16
+        assert w.dtype == torch.uint8
+        assert x.shape == (N,)
+        assert w.shape == (N, M)
+        assert rx.shape == mx.shape == (M,)
+        assert ry.shape == my.shape == (N, 1)
+        y = torch.zeros((M,), device=w.device, dtype=torch.float32)
+        torch.ops.rwkv.mm8_one(N, M, x, w, mx, rx, my, ry, y)
+        return y.to(dtype=x.dtype)
+else:
+    os.environ["RWKV_CUDA_ON"] = '0'
+
+
+@MyStatic
+def torch_mm8_seq(x, w, mx, rx, my, ry):
+    return x @ ((w.to(dtype=x.dtype) + 0.5) * ry * rx + my + mx)
+
+@MyStatic
+def torch_mm8_one(x, w, mx, rx, my, ry):
+    return x @ ((w.to(dtype=x.dtype) + 0.5) * ry * rx + my + mx)
+
+if os.environ.get('RWKV_CUDA_ON') == '1':
+    @MyStatic
+    def mm8_seq(x, w, mx, rx, my, ry):
+        if w.device.type == 'cuda' and x.dtype == torch.float16:
+            B, N, M = x.shape[0], w.shape[0], w.shape[1]
+            return cuda_mm8_seq(B, N, M, x, w, mx, rx, my, ry)
+        else:
+            return torch_mm8_seq(x, w, mx, rx, my, ry)
+    @MyStatic
+    def mm8_one(x, w, mx, rx, my, ry):
+        if w.device.type == 'cuda':
+            N, M = w.shape[0], w.shape[1]
+            return cuda_mm8_one(N, M, x, w, mx, rx, my, ry)
+        else:
+            return torch_mm8_one(x, w, mx, rx, my, ry)
+else:
+    @MyStatic
+    def mm8_seq(x, w, mx, rx, my, ry):
+        return torch_mm8_seq(x, w, mx, rx, my, ry)
+    @MyStatic
+    def mm8_one(x, w, mx, rx, my, ry):
+        return torch_mm8_one(x, w, mx, rx, my, ry)
+
+def mm8(x: torch.Tensor, w: torch.Tensor, mx: torch.Tensor, rx: torch.Tensor, my: torch.Tensor, ry: torch.Tensor):
+    if len(x.shape) == 1:
+        return mm8_one(x, w, mx, rx, my, ry)
+    return mm8_seq(x, w, mx, rx, my, ry)
+
+def matmul(a, b, mx: Optional[torch.Tensor]=None, rx: Optional[torch.Tensor]=None, my: Optional[torch.Tensor]=None, ry: Optional[torch.Tensor]=None, output_dtype: Optional[torch.dtype]=None) -> torch.Tensor:
+    if output_dtype is None:
+        output_dtype = a.dtype
+    if b.dtype in [torch.float16, torch.bfloat16, torch.float32]:
+        assert a.dtype == b.dtype
+        return matmul_float(a, b, output_dtype=output_dtype)
+    elif b.dtype == torch.uint8:
+        assert mx is not None
+        assert rx is not None
+        assert my is not None
+        assert ry is not None
+        return mm8(a, b, mx, rx, my, ry).to(output_dtype)
+    else:
+        raise ValueError("Unsupported dtype")
+
+
+if os.environ.get('RWKV_CUDA_ON') == '1' and not DISABLE_CUBLAS_GEMM:
+    def matmul_float(a, b, output_dtype: Optional[torch.dtype]=None):
+        if output_dtype is None:
+            output_dtype = a.dtype
+        if a.dtype == b.dtype == torch.float16 and a.device.type == 'cuda':
+            if len(a.shape) == 1:
+                assert len(b.shape) == 2
+                c = torch.empty((b.shape[-1],), dtype=output_dtype, device=a.device)
+                a = a.unsqueeze(0)
+            else:
+                assert len(a.shape) == len(b.shape)
+                assert len(a.shape) == 2 or len(a.shape) == 3
+                # torch.empty((*a.shape[:-1], b.shape[-1])) doesn't work with jit
+                if len(a.shape) == 2:
+                    c = torch.empty((a.shape[0], b.shape[-1]), dtype=output_dtype, device=a.device)
+                else:
+                    c = torch.empty((a.shape[0], a.shape[1], b.shape[-1]), dtype=output_dtype, device=a.device)
+            torch.ops.rwkv.gemm_fp16_cublas(a, b, c)
+            return c
+        else:
+            return (a @ b).to(output_dtype)
+
+else:
+    def matmul_float(a, b, output_dtype: Optional[torch.dtype]=None):
+        return (a @ b).to(output_dtype)
+
+
+if os.environ.get('RWKV_DML_ON') == '1':
+    import torch_directml
+    print("PyTorch with DirectML Enabled")
+
+########################################################################################################
+
+class RWKV(MyModule):
+    def __init__(self, model, strategy, verbose = True, convert_and_save_and_exit = None):
+        super().__init__()
+        if verbose:
+            prxxx = lambda *args, **kwargs: print(*args, **kwargs)
+        else:
+            prxxx = lambda *args, **kwargs: None
+
+        STRATEGY_REGEX = r"^(?:(?:^|->) *(?:cuda(?::[\d]+)?|cpu|mps|dml) (?:fp(?:16|32)|bf16)(?:i8|i4|i3)?(?: \*[\d]+\+?)? *)+$"
+        if not re.match(STRATEGY_REGEX, strategy):
+            raise ValueError("Invalid strategy. Please read https://pypi.org/project/rwkv/")
+
+        strategy = ('->'.join([x.strip() for x in strategy.split('->')])).replace('->', ' -> ')
+        self.args = types.SimpleNamespace()
+        args = self.args
+        args.MODEL_NAME = model
+        args.strategy_string = strategy
+
+        # Rescale for fp16 mode: set x = x/2 every X layer (to avoid fp16 overflow)
+        try:
+            self.RESCALE_LAYER = int(os.environ["RWKV_RESCALE_LAYER"]) # !!! NOTE: SEEMS YOU SHOULD SET IT TO 999 (disable) FOR RWKV-MUSIC MODELS !!!
+        except:
+            self.RESCALE_LAYER = 6 if 'fp16' in strategy else 0
+        prxxx(f'RWKV_JIT_ON {os.environ["RWKV_JIT_ON"]} RWKV_CUDA_ON {os.environ["RWKV_CUDA_ON"]} RESCALE_LAYER {self.RESCALE_LAYER}\n')
+
+        args.MODEL_NAME = args.MODEL_NAME.strip()
+        if not args.MODEL_NAME.endswith('.pth'):
+            args.MODEL_NAME += '.pth'
+        prxxx(f'Loading {args.MODEL_NAME} ...')
+        with torch.no_grad():
+            self.w = torch.load(args.MODEL_NAME, map_location='cpu') # load model to CPU first
+            gc.collect()
+            w = self.w
+
+            ALREADY_CONVERTED = False
+            if '_strategy' in w:
+                ALREADY_CONVERTED = True
+                assert convert_and_save_and_exit == None # you should only convert a raw model
+                prxxx(f"Converted model: strategy {w['_strategy']}, version {w['_version']}\n")
+                assert w['_strategy'] == args.strategy_string # if you are using a new strategy, re-convert the model
+                assert float(w['_version']) >= 0.7 # sometimes you should re-convert using latest convert_model.py
+                assert w['_rescale_layer'] == self.RESCALE_LAYER # must use same RESCALE_LAYER to avoid mistakes
+                del w['_strategy']
+                del w['_version']
+                del w['_rescale_layer']
+            
+            args.n_embd = w['emb.weight'].shape[1]
+            args.n_att = w['blocks.0.att.key.weight'].shape[0] # note: transposed matrix
+            args.n_ffn = w['blocks.0.ffn.key.weight'].shape[0] # note: transposed matrix
+            args.n_layer = 0
+            keys = list(w.keys())
+            self.version = 4
+            for x in keys:
+                layer_id = int(x.split('.')[1]) if ('blocks.' in x) else 0
+                args.n_layer = max(args.n_layer, layer_id+1)
+                if 'ln_x' in x:
+                    self.version = max(5, self.version)
+                if 'gate.weight' in x:
+                    self.version = max(5.1, self.version)
+                if int(self.version) == 5 and 'att.time_decay' in x:
+                    args.n_head = w[x].shape[0]
+                    if len(w[x].shape) > 1:
+                        if w[x].shape[1] > 1:
+                            self.version = max(5.2, self.version)
+                if 'time_maa' in x:
+                    self.version = max(6, self.version)
+                if int(self.version) == 6 and 'time_faaaa' in x:
+                    args.n_head = w[x].shape[0]
+            prxxx(f'Model detected: v{self.version:.1f}')
+
+            ####################### Compute strategy
+
+            s = [x.strip().split(' ') for x in strategy.split('->')]
+            plan = [0] * len(s)
+            stream_i = -1
+            stream_count = 0
+            to_allocate = args.n_layer + 1
+            allocated = 0
+            free_slots = 0
+            for i in range(len(s)):
+                si = s[i]
+                si1 = si[1]
+                if si1.startswith('fp32'): si[1] = [torch.float]
+                elif si1.startswith('fp16'): si[1] = [torch.float16]
+                elif si1.startswith('bf16'): si[1] = [torch.bfloat16]
+                if si1.endswith('i8'): si[1] += [torch.uint8]
+                else: si[1] += [si[1][0]]
+                if len(si) > 2:
+                    ss = si[2]
+                    assert ss.startswith('*')
+                    if ss.endswith('+'):
+                        plan[i] = int(ss[1:-1])
+                        stream_i = i
+                    else:
+                        plan[i] = int(ss[1:])
+                    allocated += plan[i]
+                    if allocated >= to_allocate:
+                        plan[i] += to_allocate - allocated
+                        break
+                else:
+                    free_slots += 1
+            if stream_i < 0:
+                if free_slots > 0 and to_allocate > allocated:
+                    for i in range(len(s)):
+                        if plan[i] == 0:
+                            plan[i] = (to_allocate - allocated) // free_slots
+                            allocated += plan[i]
+                            free_slots -= 1
+                if to_allocate > allocated:
+                    plan[len(s)-1] += to_allocate - allocated
+            else:
+                if to_allocate > allocated:
+                    stream_count = to_allocate - allocated
+                    plan[stream_i] += stream_count
+            prxxx(f'Strategy: (total {args.n_layer}+1={args.n_layer+1} layers)')
+            for i in range(len(s)):
+                ss = s[i]
+                if i != stream_i:
+                    prxxx(f'* {ss[0]} {str(ss[1]).replace("torch.","")}, store {plan[i]} layers')
+                else:
+                    prxxx(f'* {ss[0]} {str(ss[1]).replace("torch.","")}, store {plan[i]-stream_count} layers, stream {stream_count} layers')
+                plan[i] += (0 if i == 0 else plan[i-1])
+            self.strategy = [None] * (args.n_layer + 1)
+            strategy = self.strategy
+            for n in range(args.n_layer + 1):
+                for i in range(len(s)):
+                    if n < plan[i]:
+                        strategy[n] = types.SimpleNamespace()
+                        strategy[n].device = s[i][0]
+                        strategy[n].atype = s[i][1][0]
+                        strategy[n].wtype = s[i][1][1]
+                        strategy[n].stream = False
+                        if strategy[n].device == 'dml':
+                            strategy[n].device = torch_directml.device()
+                        if i == stream_i and n >= (plan[i] - stream_count):
+                            strategy[n].stream = True
+                        break
+                prxxx(f"{n}-{strategy[n].device}-{str(strategy[n].atype).replace('torch.','')}-{str(strategy[n].wtype).replace('torch.','')}{'-stream' if strategy[n].stream else ''}",end=' ')
+            prxxx()
+
+            ####################### Load weights to self.w
+
+            if not ALREADY_CONVERTED:
+                try: # precompute embedding
+                    w['emb.weight'] = F.layer_norm(w['emb.weight'], (args.n_embd,), weight=w['blocks.0.ln0.weight'], bias=w['blocks.0.ln0.bias'])
+                except:
+                    w['emb.weight'] = F.layer_norm(w['emb.weight'].float(), (args.n_embd,), weight=w['blocks.0.ln0.weight'].float(), bias=w['blocks.0.ln0.bias'].float())
+                # del w['blocks.0.ln0.weight']
+                # del w['blocks.0.ln0.bias']
+
+            print_need_newline = False
+
+            REAL_TIME_FIRST = False
+            for x in list(w.keys()):
+                if '.time_faaaa' in x: REAL_TIME_FIRST = True
+            if REAL_TIME_FIRST:
+                w = {k.replace('.time_faaaa','.time_first') if '.time_faaaa' in k else k: v for k, v in w.items()}
+                self.w = w
+            
+            keys = list(w.keys())
+            for x in keys:
+                w[x].requires_grad = False
+                layer_id = int(x.split('.')[1]) if ('blocks.' in x) else 0
+                if ('ln_out.' in x) or ('head.' in x):
+                    layer_id = args.n_layer
+                dd = strategy[layer_id]
+                DEVICE = dd.device
+                ATYPE = dd.atype
+                WTYPE = dd.wtype
+
+                if not ALREADY_CONVERTED:
+                    if self.RESCALE_LAYER > 0:
+                        if 'att.output.weight' in x:
+                            w[x] = w[x] / (2 ** int(layer_id // self.RESCALE_LAYER))
+                        if 'ffn.value.weight' in x:
+                            w[x] = w[x] / (2 ** int(layer_id // self.RESCALE_LAYER))
+
+                    if '.time_' in x:
+                        w[x] = w[x].squeeze()
+                    if 'key.weight' in x or 'value.weight' in x or 'receptance.weight' in x or 'gate.weight' in x or 'output.weight' in x or 'head.weight' in x:
+                        w[x] = w[x].t()
+
+                    if '.time_decay' in x and '_w' not in x: # need fp32 for this
+                        if self.version == 4:
+                            w[x] = -torch.exp(w[x].float())
+                        elif int(self.version) == 5:
+                            w[x] = torch.exp(-torch.exp(w[x].float())).reshape(-1,1,1)
+                            if self.version == 5.2:
+                                w[x] = w[x].reshape(args.n_head, -1, 1)
+                        elif self.version == 6.0:
+                            w[x] = w[x].float().reshape(args.n_head, -1, 1)
+                    elif '.time_first' in x: # need fp32 for this
+                        if self.version == 4:
+                            w[x] = w[x].float()
+                        elif int(self.version) in [5, 6]:
+                            if REAL_TIME_FIRST:
+                                w[x] = w[x].float().reshape(-1,1,1)
+                            else:
+                                w[x] = torch.exp(w[x].float()).reshape(-1,1,1)
+                            if self.version in [5.2, 6.0]:
+                                w[x] = w[x].reshape(args.n_head, -1, 1)
+                    elif '.ln_x' in x: # need fp32 for group_norm
+                        w[x] = w[x].float()
+                    else:
+                        if (len(w[x].shape) == 2) and ('emb' not in x):
+                            if WTYPE != torch.uint8:
+                                w[x] = w[x].to(dtype=WTYPE)
+                            else:
+                                w[x] = w[x].float()
+
+                                if w[x].shape[0] > w[x].shape[1]:
+                                    w[x+'_my'] = torch.amin(w[x], dim=1).unsqueeze(1)
+                                    w[x] = w[x] - w[x+'_my']
+                                    w[x+'_mx'] = torch.amin(w[x], dim=0)
+                                    w[x] = w[x] - w[x+'_mx']
+                                    w[x+'_rx'] = torch.amax(w[x], dim=0)
+                                    w[x] = w[x] / w[x+'_rx']
+                                    w[x+'_ry'] = torch.amax(w[x], dim=1).unsqueeze(1)
+                                    w[x] = w[x] / w[x+'_ry']
+                                else:
+                                    w[x+'_mx'] = torch.amin(w[x], dim=0)
+                                    w[x] = w[x] - w[x+'_mx']
+                                    w[x+'_my'] = torch.amin(w[x], dim=1).unsqueeze(1)
+                                    w[x] = w[x] - w[x+'_my']
+                                    w[x+'_rx'] = torch.amax(w[x], dim=0)
+                                    w[x] = w[x] / w[x+'_rx']
+                                    w[x+'_ry'] = torch.amax(w[x], dim=1).unsqueeze(1)
+                                    w[x] = w[x] / w[x+'_ry']
+
+                                w[x] = torch.clip(torch.floor(w[x] * 256), min=0, max=255).to(dtype=torch.uint8)
+                                w[x+'_mx'] = w[x+'_mx'].to(dtype=ATYPE).contiguous()
+                                w[x+'_rx'] = (w[x+'_rx'] / 16).to(dtype=ATYPE).contiguous()
+                                w[x+'_my'] = w[x+'_my'].to(dtype=ATYPE).contiguous()
+                                w[x+'_ry'] = (w[x+'_ry'] / 16).to(dtype=ATYPE).contiguous()
+                        else:
+                            w[x] = w[x].to(dtype=ATYPE)
+                
+                if convert_and_save_and_exit == None:
+                    if 'emb.' in x:
+                        w[x] = w[x].contiguous()
+                    elif (dd.stream) and (x.endswith('key.weight') or x.endswith('value.weight') or x.endswith('receptance.weight') or x.endswith('output.weight')):
+                        try:
+                            w[x] = w[x].contiguous().pin_memory() # if you see "CUDA error: out of memory" here, that's out of CPU RAM, not VRAM. Get more RAM :)
+                        except:
+                            print('Note: You are running out of RAM. Get more CPU RAM. Now this will run much slower.')
+                    elif DEVICE != 'cpu':
+                        w[x] = w[x].to(device=DEVICE).contiguous()
+                    
+                    if (dd.stream) or (DEVICE != 'cpu'):
+                        try:
+                            w[x+'_mx'] = w[x+'_mx'].to(device=DEVICE).contiguous()
+                            w[x+'_rx'] = w[x+'_rx'].to(device=DEVICE).contiguous()
+                            w[x+'_my'] = w[x+'_my'].to(device=DEVICE).contiguous()
+                            w[x+'_ry'] = w[x+'_ry'].to(device=DEVICE).contiguous()
+                        except:
+                            pass
+
+                if 'ffn.value.weight' in x:
+                    gc.collect()
+                    if 'cuda' in args.strategy_string:
+                        torch.cuda.empty_cache()
+
+                shape = [i for i in w[x].shape if i != 1]
+                if len(shape) > 1:
+                    shape = f" {str(shape[0]).rjust(5)} {str(shape[1]).rjust(5)}"
+                else:
+                    shape = f" {str(shape[0]).rjust(5)}      "
+                if layer_id == 0 or layer_id >= args.n_layer-1:
+                    if print_need_newline:
+                        prxxx('\n', end = '')
+                        print_need_newline = False
+                    dt = str(w[x].dtype).replace('torch.', '')
+                    dt = dt.replace('float32', 'f32').replace('bfloat16', 'bf16').replace('float16', 'f16').replace('uint8', 'i8')
+                    prxxx(x.ljust(32), dt.rjust(4), str(w[x].device).rjust(8), shape, ' (pinned)' if w[x].is_pinned() else '')
+                else:
+                    print_need_newline = True
+                    prxxx('.', end = '', flush = True)
+            
+            if convert_and_save_and_exit:
+                w['_strategy'] = args.strategy_string
+                w['_rescale_layer'] = self.RESCALE_LAYER
+                w['_version'] = '0.7'
+                if not convert_and_save_and_exit.endswith('.pth'):
+                    convert_and_save_and_exit += '.pth'
+                prxxx(f'Saving to {convert_and_save_and_exit}...')
+                torch.save(w, convert_and_save_and_exit)
+                prxxx(f'Converted and saved. Now this will exit.')
+                exit(0)
+            
+            if self.version == 5.2 and os.environ["RWKV_CUDA_ON"] == '1':
+                HEAD_SIZE = args.n_att // args.n_head
+                rwkv5 = load(name="rwkv5", sources=[f"{current_path}/cuda/rwkv5_op.cpp", f"{current_path}/cuda/rwkv5.cu"],
+                                verbose=True, extra_cuda_cflags=["-res-usage", "--use_fast_math", "-O3", "-Xptxas -O3" if os.name != "nt" else "", "--extra-device-vectorization", f"-D_N_={HEAD_SIZE}"])
+
+                class RWKV_5(torch.autograd.Function):
+                    @staticmethod
+                    def forward(ctx, B, T, C, H, state, r, k, v, w, u):
+                        with torch.no_grad():
+                            assert HEAD_SIZE == C // H
+                            ctx.B = B
+                            ctx.T = T
+                            ctx.C = C
+                            ctx.H = H
+                            assert state.dtype == torch.float32
+                            assert w.dtype == torch.float32
+                            assert r.is_contiguous()
+                            assert k.is_contiguous()
+                            assert v.is_contiguous()
+                            assert w.is_contiguous()                            
+                            assert u.is_contiguous()                            
+                            assert state.is_contiguous()
+
+                            y = torch.empty((B, T, C), device=w.device, dtype=r.dtype, memory_format=torch.contiguous_format)
+                            if r.dtype == torch.bfloat16:
+                                rwkv5.forward_bf16(B, T, C, H, state, r, k, v, w, u, y)
+                            elif r.dtype == torch.float16:
+                                rwkv5.forward_fp16(B, T, C, H, state, r, k, v, w, u, y)
+                            elif r.dtype == torch.float32:
+                                rwkv5.forward_fp32(B, T, C, H, state, r, k, v, w, u, y)
+                            return y, state
+                self.RWKV_5 = RWKV_5
+
+            if self.version == 6.0 and os.environ["RWKV_CUDA_ON"] == '1':
+                HEAD_SIZE = args.n_att // args.n_head
+                rwkv6 = load(name="rwkv6", sources=[f"{current_path}/cuda/rwkv6_op.cpp", f"{current_path}/cuda/rwkv6.cu"],
+                                verbose=True, extra_cuda_cflags=["-res-usage", "--use_fast_math", "-O3", "-Xptxas -O3", "--extra-device-vectorization", f"-D_N_={HEAD_SIZE}", f"-D_T_={4096}"])
+                    
+                class RWKV_6(torch.autograd.Function):
+                    @staticmethod
+                    def forward(ctx, B, T, C, H, state, r, k, v, w, u):
+                        with torch.no_grad():
+                            assert HEAD_SIZE == C // H
+                            ctx.B = B
+                            ctx.T = T
+                            ctx.C = C
+                            ctx.H = H
+                            assert state.dtype == torch.float32
+                            assert w.dtype == torch.float32
+                            assert r.is_contiguous()
+                            assert k.is_contiguous()
+                            assert v.is_contiguous()
+                            assert w.is_contiguous()
+                            assert u.is_contiguous()
+                            eew = torch.exp(-torch.exp(w.float())).contiguous()
+
+                            y = torch.empty((B, T, C), device=w.device, dtype=r.dtype, memory_format=torch.contiguous_format)
+                            if r.dtype == torch.bfloat16:
+                                rwkv6.forward_bf16(B, T, C, H, state, r, k, v, eew, u, y)
+                            elif r.dtype == torch.float16:
+                                rwkv6.forward_fp16(B, T, C, H, state, r, k, v, eew, u, y)
+                            elif r.dtype == torch.float32:
+                                rwkv6.forward_fp32(B, T, C, H, state, r, k, v, eew, u, y)
+                            return y, state
+                self.RWKV_6 = RWKV_6
+        
+            gc.collect()
+            if 'cuda' in args.strategy_string:
+                torch.cuda.empty_cache()
+
+    def RUN_RWKV_5(self, B, T, C, H, state, r, k, v, w, u):
+        return self.RWKV_5.apply(B, T, C, H, state, r, k, v, w, u)
+
+    def RUN_RWKV_6(self, B, T, C, H, state, r, k, v, w, u):
+        return self.RWKV_6.apply(B, T, C, H, state, r, k, v, w, u)
+
+    ########################################################################################################
+
+    @MyFunction
+    def ffn_one(self, x, sx, ln_w, ln_b, k_mix, r_mix, kw, vw, rw, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry):
+        xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
+        kx = xx * k_mix + sx * (1 - k_mix)
+        rx = xx * r_mix + sx * (1 - r_mix)
+
+        r = torch.sigmoid(matmul(rx, rw, rmx, rrx, rmy, rry))
+        vx = torch.relu(matmul(kx, kw, kmx, krx, kmy, kry)) ** 2
+        out = r * matmul(vx, vw, vmx, vrx, vmy, vry)
+        return x + out, xx
+
+    @MyFunction
+    def ffn_seq(self, x, sx, ln_w, ln_b, k_mix, r_mix, kw, vw, rw, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry):
+        xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
+        sx = torch.cat((sx.unsqueeze(0), xx[:-1,:]))
+        kx = xx * k_mix + sx * (1 - k_mix)
+        rx = xx * r_mix + sx * (1 - r_mix)
+
+        r = torch.sigmoid(matmul(rx, rw, rmx, rrx, rmy, rry))
+        vx = torch.relu(matmul(kx, kw, kmx, krx, kmy, kry)) ** 2
+        out = r * matmul(vx, vw, vmx, vrx, vmy, vry)
+        return x + out, xx[-1,:]
+
+    @MyFunction
+    def ffn_one_v6(self, x, sx, ln_w, ln_b, k_maa, r_maa, kw, vw, rw, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry):
+        xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
+        sx = sx - xx
+        kx = xx + sx * k_maa
+        rx = xx + sx * r_maa
+
+        r = torch.sigmoid(matmul(rx, rw, rmx, rrx, rmy, rry))
+        vx = torch.relu(matmul(kx, kw, kmx, krx, kmy, kry)) ** 2
+        out = r * matmul(vx, vw, vmx, vrx, vmy, vry)
+        return x + out, xx
+
+    @MyFunction
+    def ffn_seq_v6(self, x, sx, ln_w, ln_b, k_maa, r_maa, kw, vw, rw, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry):
+        xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
+        sx = torch.cat((sx.unsqueeze(0), xx[:-1,:]))
+        sx = sx - xx
+        kx = xx + sx * k_maa
+        rx = xx + sx * r_maa
+
+        r = torch.sigmoid(matmul(rx, rw, rmx, rrx, rmy, rry))
+        vx = torch.relu(matmul(kx, kw, kmx, krx, kmy, kry)) ** 2
+        out = r * matmul(vx, vw, vmx, vrx, vmy, vry)
+        return x + out, xx[-1,:]
+
+    ########################################################################################################
+
+    @MyFunction
+    def att_one(self, x, sx, aa, bb, pp, ln_w, ln_b, k_mix, v_mix, r_mix, t_decay, t_first, kw, vw, rw, ow, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry, omx, orx, omy, ory):
+        xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
+        kx = xx * k_mix + sx * (1 - k_mix)
+        vx = xx * v_mix + sx * (1 - v_mix)
+        rx = xx * r_mix + sx * (1 - r_mix)
+
+        r = torch.sigmoid(matmul(rx, rw, rmx, rrx, rmy, rry))
+        k = matmul(kx, kw, kmx, krx, kmy, kry, output_dtype=torch.float32)
+        v = matmul(vx, vw, vmx, vrx, vmy, vry, output_dtype=torch.float32)
+
+        ww = t_first + k
+        p = torch.maximum(pp, ww)
+        e1 = torch.exp(pp - p)
+        e2 = torch.exp(ww - p)
+        wkv = ((e1 * aa + e2 * v) / (e1 * bb + e2)).to(dtype=x.dtype)
+        ww = t_decay + pp
+        p = torch.maximum(ww, k)
+        e1 = torch.exp(ww - p)
+        e2 = torch.exp(k - p)
+
+        out = matmul(r * wkv, ow, omx, orx, omy, ory)
+        return x + out, xx, e1 * aa + e2 * v, e1 * bb + e2, p
+
+    @MyFunction
+    def att_seq(self, x, sx, aa, bb, pp, ln_w, ln_b, k_mix, v_mix, r_mix, t_decay, t_first, kw, vw, rw, ow, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry, omx, orx, omy, ory):
+        xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
+        sx = torch.cat((sx.unsqueeze(0), xx[:-1,:]))
+        kx = xx * k_mix + sx * (1 - k_mix)
+        vx = xx * v_mix + sx * (1 - v_mix)
+        rx = xx * r_mix + sx * (1 - r_mix)
+
+        r = torch.sigmoid(matmul(rx, rw, rmx, rrx, rmy, rry))
+        k = matmul(kx, kw, kmx, krx, kmy, kry, output_dtype=torch.float32)
+        v = matmul(vx, vw, vmx, vrx, vmy, vry, output_dtype=torch.float32)
+
+        T = x.shape[0]
+        for t in range(T):
+            kk = k[t]
+            vv = v[t]
+            ww = t_first + kk
+            p = torch.maximum(pp, ww)
+            e1 = torch.exp(pp - p)
+            e2 = torch.exp(ww - p)
+            sx[t] = ((e1 * aa + e2 * vv) / (e1 * bb + e2)).to(dtype=x.dtype)
+            ww = t_decay + pp
+            p = torch.maximum(ww, kk)
+            e1 = torch.exp(ww - p)
+            e2 = torch.exp(kk - p)
+            aa = e1 * aa + e2 * vv
+            bb = e1 * bb + e2
+            pp = p
+        out = matmul(r * sx, ow, omx, orx, omy, ory)
+        return x + out, xx[-1,:], aa, bb, pp
+
+    ########################################################################################################
+
+    @MyFunction
+    def att_one_v5(self, x, sx, s, ln_w, ln_b, lx_w, lx_b, k_mix, v_mix, r_mix, t_decay, t_first, kw, vw, rw, ow, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry, omx, orx, omy, ory):
+        xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
+        kx = xx * k_mix + sx * (1 - k_mix)
+        vx = xx * v_mix + sx * (1 - v_mix)
+        rx = xx * r_mix + sx * (1 - r_mix)
+
+        H = t_decay.shape[0]
+        N = x.shape[-1] // H
+
+        r = matmul(rx, rw, rmx, rrx, rmy, rry, output_dtype=torch.float32).view(H, 1, N)
+        k = matmul(kx, kw, kmx, krx, kmy, kry, output_dtype=torch.float32).view(H, N, 1)
+        v = matmul(vx, vw, vmx, vrx, vmy, vry, output_dtype=torch.float32).view(H, 1, N)
+        
+        a = matmul(k, v)
+        out = r @ (t_first * a + s)
+        s = a + t_decay * s
+
+        out = out.flatten()
+        out = F.group_norm(out.unsqueeze(0), num_groups=H, weight=lx_w, bias=lx_b, eps = 64e-5).squeeze(0)
+        out = out.to(dtype=x.dtype)
+        out = matmul(out, ow, omx, orx, omy, ory)
+
+        return x + out, xx, s
+
+    @MyFunction
+    def att_seq_v5(self, x, sx, s, ln_w, ln_b, lx_w, lx_b, k_mix, v_mix, r_mix, t_decay, t_first, kw, vw, rw, ow, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry, omx, orx, omy, ory):
+        xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
+        sx = torch.cat((sx.unsqueeze(0), xx[:-1,:]))
+        kx = xx * k_mix + sx * (1 - k_mix)
+        vx = xx * v_mix + sx * (1 - v_mix)
+        rx = xx * r_mix + sx * (1 - r_mix)
+
+        H = t_decay.shape[0]
+        N = x.shape[-1] // H
+        T = x.shape[0]
+
+        w = t_decay.reshape(-1, 1)
+        u = t_first.reshape(-1, 1)
+        ws = w.pow(T).reshape(H, 1, 1)
+        ind = torch.arange(T-1, -1, -1, device=w.device).unsqueeze(0).repeat(H, 1)
+        w = w.repeat(1, T).pow(ind)
+        wk = w.reshape(H, 1, T)
+        wb = wk.transpose(-2, -1).flip(1)
+        w = torch.cat([w[:, 1:], u], dim=1)
+        w = F.pad(w, (0, T))
+        w = torch.tile(w, [T])
+        w = w[:, :-T].reshape(-1, T, 2 * T - 1)
+        w = w[:, :, T-1:].reshape(H, T, T)
+
+        r = matmul(rx, rw, rmx, rrx, rmy, rry, output_dtype=torch.float32).view(T, H, N).transpose(0, 1)
+        k = matmul(kx, kw, kmx, krx, kmy, kry, output_dtype=torch.float32).view(T, H, N).permute(1, 2, 0)
+        v = matmul(vx, vw, vmx, vrx, vmy, vry, output_dtype=torch.float32).view(T, H, N).transpose(0, 1)
+
+        out = ((r @ k) * w) @ v + (r @ s) * wb
+        s = ws * s + (k * wk) @ v
+        
+        out = out.transpose(0, 1).contiguous().reshape(T, H*N)
+        out = F.group_norm(out, num_groups=H, weight=lx_w, bias=lx_b, eps = 64e-5)
+        out = out.to(dtype=x.dtype)
+        out = matmul(out, ow, omx, orx, omy, ory)
+
+        return x + out, xx[-1,:], s
+
+    ########################################################################################################
+
+    @MyFunction
+    def att_one_v5_1(self, x, sx, s, ln_w, ln_b, lx_w, lx_b, k_mix, v_mix, r_mix, g_mix, t_decay, t_first, kw, vw, rw, gw, ow, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry, gmx, grx, gmy, gry, omx, orx, omy, ory):
+        xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
+        kx = xx * k_mix + sx * (1 - k_mix)
+        vx = xx * v_mix + sx * (1 - v_mix)
+        rx = xx * r_mix + sx * (1 - r_mix)
+        gx = xx * g_mix + sx * (1 - g_mix)
+
+        H = t_decay.shape[0]
+        N = x.shape[-1] // H
+
+        r = matmul(rx, rw, rmx, rrx, rmy, rry, output_dtype=torch.float32).view(H, 1, N)
+        k = matmul(kx, kw, kmx, krx, kmy, kry, output_dtype=torch.float32).view(H, N, 1)
+        v = matmul(vx, vw, vmx, vrx, vmy, vry, output_dtype=torch.float32).view(H, 1, N)
+        g = F.silu(matmul(gx, gw, gmx, grx, gmy, gry))
+        
+        a = matmul(k, v)
+        out = r @ (t_first * a + s)
+        s = a + t_decay * s
+
+        out = out.flatten()
+        out = F.group_norm(out.unsqueeze(0), num_groups=H, weight=lx_w, bias=lx_b, eps = 64e-5).squeeze(0)
+        out = out.to(dtype=x.dtype) * g
+        out = matmul(out, ow, omx, orx, omy, ory)
+
+        return x + out, xx, s
+
+    @MyFunction
+    def att_seq_v5_1(self, x, sx, s, ln_w, ln_b, lx_w, lx_b, k_mix, v_mix, r_mix, g_mix, t_decay, t_first, kw, vw, rw, gw, ow, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry, gmx, grx, gmy, gry, omx, orx, omy, ory):
+        xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
+        sx = torch.cat((sx.unsqueeze(0), xx[:-1,:]))
+        kx = xx * k_mix + sx * (1 - k_mix)
+        vx = xx * v_mix + sx * (1 - v_mix)
+        rx = xx * r_mix + sx * (1 - r_mix)
+        gx = xx * g_mix + sx * (1 - g_mix)
+
+        H = t_decay.shape[0]
+        N = x.shape[-1] // H
+        T = x.shape[0]
+
+        w = t_decay.reshape(-1, 1)
+        u = t_first.reshape(-1, 1)
+        ws = w.pow(T).reshape(H, 1, 1)
+        ind = torch.arange(T-1, -1, -1, device=w.device).unsqueeze(0).repeat(H, 1)
+        w = w.repeat(1, T).pow(ind)
+        wk = w.reshape(H, 1, T)
+        wb = wk.transpose(-2, -1).flip(1)
+        w = torch.cat([w[:, 1:], u], dim=1)
+        w = F.pad(w, (0, T))
+        w = torch.tile(w, [T])
+        w = w[:, :-T].reshape(-1, T, 2 * T - 1)
+        w = w[:, :, T-1:].reshape(H, T, T)
+
+        r = matmul(rx, rw, rmx, rrx, rmy, rry, output_dtype=torch.float32).view(T, H, N).transpose(0, 1)
+        k = matmul(kx, kw, kmx, krx, kmy, kry, output_dtype=torch.float32).view(T, H, N).permute(1, 2, 0)
+        v = matmul(vx, vw, vmx, vrx, vmy, vry, output_dtype=torch.float32).view(T, H, N).transpose(0, 1)
+        g = F.silu(matmul(gx, gw, gmx, grx, gmy, gry))
+
+        out = ((r @ k) * w) @ v + (r @ s) * wb
+        s = ws * s + (k * wk) @ v
+        
+        out = out.transpose(0, 1).contiguous().reshape(T, H*N)
+        out = F.group_norm(out, num_groups=H, weight=lx_w, bias=lx_b, eps = 64e-5)
+        out = out.to(dtype=x.dtype) * g
+        out = matmul(out, ow, omx, orx, omy, ory)
+
+        return x + out, xx[-1,:], s
+
+    ########################################################################################################
+
+    @MyFunction
+    def att_seq_v5_2(self, x, sx, s, ln_w, ln_b, lx_w, lx_b, k_mix, v_mix, r_mix, g_mix, t_decay, t_first, kw, vw, rw, gw, ow, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry, gmx, grx, gmy, gry, omx, orx, omy, ory):
+        xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
+        sx = torch.cat((sx.unsqueeze(0), xx[:-1,:]))
+        kx = xx * k_mix + sx * (1 - k_mix)
+        vx = xx * v_mix + sx * (1 - v_mix)
+        rx = xx * r_mix + sx * (1 - r_mix)
+        gx = xx * g_mix + sx * (1 - g_mix)
+
+        H = t_decay.shape[0]
+        N = x.shape[-1] // H
+        T = x.shape[0]
+
+        r = matmul(rx, rw, rmx, rrx, rmy, rry, output_dtype=torch.float32).view(T, H, N).transpose(0, 1)
+        k = matmul(kx, kw, kmx, krx, kmy, kry, output_dtype=torch.float32).view(T, H, N).permute(1, 2, 0)
+        v = matmul(vx, vw, vmx, vrx, vmy, vry, output_dtype=torch.float32).view(T, H, N).transpose(0, 1)
+        g = F.silu(matmul(gx, gw, gmx, grx, gmy, gry))
+
+        out = torch.empty((T, H, N), dtype=r.dtype, device=r.device)
+        for t in range(T):
+            rt = r[:,t:t+1,:]
+            kt = k[:,:,t:t+1]
+            vt = v[:,t:t+1,:]
+            at = matmul(kt, vt)
+            out[t] = (rt @ (t_first * at + s)).squeeze(1)
+            s = at + t_decay * s
+
+        out = out.reshape(T, H*N)
+        out = F.group_norm(out, num_groups=H, weight=lx_w, bias=lx_b, eps = 64e-5)
+        out = out.to(dtype=x.dtype) * g
+        out = matmul(out, ow, omx, orx, omy, ory)
+
+        return x + out, xx[-1,:], s
+
+    ########################################################################################################
+
+    @MyFunction
+    def att_one_v6_0(self, x, sx, s, ln_w, ln_b, lx_w, lx_b, x_maa, w_maa, k_maa, v_maa, r_maa, g_maa, tm_w1, tm_w2, td_w1, td_w2, t_decay, t_first, kw, vw, rw, gw, ow, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry, gmx, grx, gmy, gry, omx, orx, omy, ory):
+        xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
+        
+        sx = sx - xx
+        xxx = xx + sx * x_maa
+        xxx = torch.tanh(xxx @ tm_w1).view(5, 1, -1)
+        xxx = torch.bmm(xxx, tm_w2).view(5, -1)
+        mw, mk, mv, mr, mg = xxx.unbind(dim=0)
+
+        wx = xx + sx * (w_maa + mw)
+        kx = xx + sx * (k_maa + mk)
+        vx = xx + sx * (v_maa + mv)
+        rx = xx + sx * (r_maa + mr)
+        gx = xx + sx * (g_maa + mg)
+
+        H = t_decay.shape[0]
+        N = x.shape[-1] // H
+
+        r = matmul(rx, rw, rmx, rrx, rmy, rry, output_dtype=torch.float32).view(H, 1, N)
+        k = matmul(kx, kw, kmx, krx, kmy, kry, output_dtype=torch.float32).view(H, N, 1)
+        v = matmul(vx, vw, vmx, vrx, vmy, vry, output_dtype=torch.float32).view(H, 1, N)
+        g = F.silu(matmul(gx, gw, gmx, grx, gmy, gry))
+        
+        w = t_decay + (torch.tanh(wx @ td_w1) @ td_w2).float().view(H, N, 1)
+        w = torch.exp(-torch.exp(w.float()))
+
+        a = matmul(k, v)
+        out = r @ (t_first * a + s)
+        s = a + w * s
+
+        out = out.flatten()
+        out = F.group_norm(out.unsqueeze(0), num_groups=H, weight=lx_w, bias=lx_b, eps = 64e-5).squeeze(0)
+        out = out.to(dtype=x.dtype) * g
+        out = matmul(out, ow, omx, orx, omy, ory)
+
+        return x + out, xx, s
+
+    @MyFunction
+    def att_seq_v6_0(self, x, sx, s, ln_w, ln_b, lx_w, lx_b, x_maa, w_maa, k_maa, v_maa, r_maa, g_maa, tm_w1, tm_w2, td_w1, td_w2, t_decay, t_first, kw, vw, rw, gw, ow, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry, gmx, grx, gmy, gry, omx, orx, omy, ory):
+        H = t_decay.shape[0]
+        N = x.shape[-1] // H
+        T = x.shape[0]
+
+        xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
+        sx = torch.cat((sx.unsqueeze(0), xx[:-1,:])) - xx
+        xxx = xx + sx * x_maa
+        xxx = torch.tanh(xxx @ tm_w1).view(T, 5, -1).transpose(0, 1)
+        xxx = torch.bmm(xxx, tm_w2).view(5, T, -1)
+        mw, mk, mv, mr, mg = xxx.unbind(dim=0)
+
+        wx = xx + sx * (w_maa + mw)
+        kx = xx + sx * (k_maa + mk)
+        vx = xx + sx * (v_maa + mv)
+        rx = xx + sx * (r_maa + mr)
+        gx = xx + sx * (g_maa + mg)
+
+        r = matmul(rx, rw, rmx, rrx, rmy, rry, output_dtype=torch.float32).view(T, H, N).transpose(0, 1)
+        k = matmul(kx, kw, kmx, krx, kmy, kry, output_dtype=torch.float32).view(T, H, N).permute(1, 2, 0)
+        v = matmul(vx, vw, vmx, vrx, vmy, vry, output_dtype=torch.float32).view(T, H, N).transpose(0, 1)
+        g = F.silu(matmul(gx, gw, gmx, grx, gmy, gry))
+
+        w = t_decay.view(1, H, N, 1) + (torch.tanh(wx @ td_w1) @ td_w2).float().view(T, H, N, 1)
+        w = torch.exp(-torch.exp(w.float()))
+        out = torch.empty((T, H, N), dtype=r.dtype, device=r.device)
+        for t in range(T):
+            rt = r[:,t:t+1,:]
+            kt = k[:,:,t:t+1]
+            vt = v[:,t:t+1,:]
+            at = matmul(kt, vt)
+            out[t] = (rt @ (t_first * at + s)).squeeze(1)
+            s = at + w[t] * s
+
+        out = out.reshape(T, H*N)
+        out = F.group_norm(out, num_groups=H, weight=lx_w, bias=lx_b, eps = 64e-5)
+        out = out.to(dtype=x.dtype) * g
+        out = matmul(out, ow, omx, orx, omy, ory)
+
+        return x + out, xx[-1,:], s
+
+    ########################################################################################################
+
+    if os.environ["RWKV_CUDA_ON"] == '1':
+        @MyFunction
+        def cuda_att_seq(self, x, sx, aa, bb, pp, ln_w, ln_b, k_mix, v_mix, r_mix, t_decay, t_first, kw, vw, rw, ow, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry, omx, orx, omy, ory):
+            T, C = x.shape
+            xx = F.layer_norm(x, (C,), weight=ln_w, bias=ln_b)
+            sx = torch.cat((sx.unsqueeze(0), xx[:-1,:]))
+            kx = xx * k_mix + sx * (1 - k_mix)
+            vx = xx * v_mix + sx * (1 - v_mix)
+            rx = xx * r_mix + sx * (1 - r_mix)
+
+            r = torch.sigmoid(matmul(rx, rw, rmx, rrx, rmy, rry))
+            k = matmul(kx, kw, kmx, krx, kmy, kry, output_dtype=torch.float32)
+            v = matmul(vx, vw, vmx, vrx, vmy, vry, output_dtype=torch.float32)
+            y, aa, bb, pp = cuda_wkv(T, C, t_decay, t_first, k, v, aa, bb, pp)
+
+            out = matmul(r * y.to(x.dtype), ow, omx, orx, omy, ory)
+            return x + out, xx[-1,:], aa, bb, pp
+
+        @MyFunction
+        def v5_2_before(self, x, sx, s, ln_w, ln_b, lx_w, lx_b, k_mix, v_mix, r_mix, g_mix, t_decay, t_first, kw, vw, rw, gw, ow, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry, gmx, grx, gmy, gry, omx, orx, omy, ory):
+            xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
+            sx = torch.cat((sx.unsqueeze(0), xx[:-1,:]))
+            kx = xx * k_mix + sx * (1 - k_mix)
+            vx = xx * v_mix + sx * (1 - v_mix)
+            rx = xx * r_mix + sx * (1 - r_mix)
+            gx = xx * g_mix + sx * (1 - g_mix)
+
+            r = matmul(rx, rw, rmx, rrx, rmy, rry, output_dtype=torch.float32)
+            k = matmul(kx, kw, kmx, krx, kmy, kry, output_dtype=torch.float32)
+            v = matmul(vx, vw, vmx, vrx, vmy, vry, output_dtype=torch.float32)
+            g = F.silu(matmul(gx, gw, gmx, grx, gmy, gry))
+
+            return r, k, v, g, xx[-1,:], s.transpose(-1,-2).contiguous()
+
+        @MyFunction
+        def v5_2_after(self, t_decay, out, s, x, xxx, g, lx_w, lx_b, ow, omx, orx, omy, ory):
+            H = t_decay.shape[0]
+            N = x.shape[-1] // H
+            T = x.shape[0]
+
+            s = s.transpose(-1,-2)
+            out = out.reshape(T, H*N)
+            out = F.group_norm(out, num_groups=H, weight=lx_w, bias=lx_b, eps = 64e-5)
+            out = out.to(dtype=x.dtype) * g
+            out = matmul(out, ow, omx, orx, omy, ory)
+
+            return x + out, xxx, s
+
+        def cuda_att_seq_v5_2(self, x, sx, s, ln_w, ln_b, lx_w, lx_b, k_mix, v_mix, r_mix, g_mix, t_decay, t_first, kw, vw, rw, gw, ow, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry, gmx, grx, gmy, gry, omx, orx, omy, ory):
+            H = t_decay.shape[0]
+            N = x.shape[-1] // H
+            T = x.shape[0]
+
+            r, k, v, g, xxx, ss = self.v5_2_before(x, sx, s, ln_w, ln_b, lx_w, lx_b, k_mix, v_mix, r_mix, g_mix, t_decay, t_first, kw, vw, rw, gw, ow, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry, gmx, grx, gmy, gry, omx, orx, omy, ory)
+
+            out, s = self.RUN_RWKV_5(1, T, self.args.n_att, H, ss, r, k, v, w=t_decay, u=t_first)
+
+            return self.v5_2_after(t_decay, out, s, x, xxx, g, lx_w, lx_b, ow, omx, orx, omy, ory)
+
+        @MyFunction
+        def v6_0_before(self, x, sx, s, ln_w, ln_b, lx_w, lx_b, x_maa, w_maa, k_maa, v_maa, r_maa, g_maa, tm_w1, tm_w2, td_w1, td_w2, t_decay, t_first, kw, vw, rw, gw, ow, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry, gmx, grx, gmy, gry, omx, orx, omy, ory):
+            H = t_decay.shape[0]
+            N = x.shape[-1] // H
+            T = x.shape[0]
+
+            xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
+            sx = torch.cat((sx.unsqueeze(0), xx[:-1,:])) - xx
+            xxx = xx + sx * x_maa
+            xxx = torch.tanh(xxx @ tm_w1).view(T, 5, -1).transpose(0, 1)
+            xxx = torch.bmm(xxx, tm_w2).view(5, T, -1)
+            mw, mk, mv, mr, mg = xxx.unbind(dim=0)
+
+            wx = xx + sx * (w_maa + mw)
+            kx = xx + sx * (k_maa + mk)
+            vx = xx + sx * (v_maa + mv)
+            rx = xx + sx * (r_maa + mr)
+            gx = xx + sx * (g_maa + mg)
+
+            r = matmul(rx, rw, rmx, rrx, rmy, rry, output_dtype=torch.float32)
+            k = matmul(kx, kw, kmx, krx, kmy, kry, output_dtype=torch.float32)
+            v = matmul(vx, vw, vmx, vrx, vmy, vry, output_dtype=torch.float32)
+            g = F.silu(matmul(gx, gw, gmx, grx, gmy, gry))
+
+            w = t_decay.view(1, H, N, 1) + (torch.tanh(wx @ td_w1) @ td_w2).float().view(T, H, N, 1)
+
+            return r, k, v, g, w, xx[-1,:], s.transpose(-1,-2).contiguous()
+
+        def cuda_att_seq_v6_0(self, x, sx, s, ln_w, ln_b, lx_w, lx_b, x_maa, w_maa, k_maa, v_maa, r_maa, g_maa, tm_w1, tm_w2, td_w1, td_w2, t_decay, t_first, kw, vw, rw, gw, ow, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry, gmx, grx, gmy, gry, omx, orx, omy, ory):
+            H = t_decay.shape[0]
+            N = x.shape[-1] // H
+            T = x.shape[0]
+
+            r, k, v, g, w, xxx, ss = self.v6_0_before(x, sx, s, ln_w, ln_b, lx_w, lx_b, x_maa, w_maa, k_maa, v_maa, r_maa, g_maa, tm_w1, tm_w2, td_w1, td_w2, t_decay, t_first, kw, vw, rw, gw, ow, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry, gmx, grx, gmy, gry, omx, orx, omy, ory)
+
+            out, s = self.RUN_RWKV_6(1, T, self.args.n_att, H, ss, r, k, v, w=w, u=t_first)
+            return self.v5_2_after(t_decay, out, s, x, xxx, g, lx_w, lx_b, ow, omx, orx, omy, ory)
+
+    ########################################################################################################
+
+    def forward(self, tokens=None, state=None, full_output=False, embs=None):
+        with torch.no_grad():
+            w = self.w
+            args = self.args
+
+            if state == None:
+                if self.version == 4:
+                    state = [None] * args.n_layer * 5
+                    for i in range(args.n_layer): # state: 0=att_xx 1=att_aa 2=att_bb 3=att_pp 4=ffn_xx
+                        dd = self.strategy[i]
+                        dev = dd.device
+                        atype = dd.atype
+                        state[i*5+0] = torch.zeros(args.n_embd, dtype=atype, requires_grad=False, device=dev).contiguous()
+                        state[i*5+1] = torch.zeros(args.n_att, dtype=torch.float, requires_grad=False, device=dev).contiguous()
+                        state[i*5+2] = torch.zeros(args.n_att, dtype=torch.float, requires_grad=False, device=dev).contiguous()
+                        state[i*5+3] = torch.zeros(args.n_att, dtype=torch.float, requires_grad=False, device=dev).contiguous() - 1e30
+                        state[i*5+4] = torch.zeros(args.n_embd, dtype=atype, requires_grad=False, device=dev).contiguous()
+                elif int(self.version) in [5,6]:
+                    state = [None] * args.n_layer * 3
+                    for i in range(args.n_layer): # state: 0=att_xx 1=att_kv 2=ffn_xx
+                        dd = self.strategy[i]
+                        dev = dd.device
+                        atype = dd.atype
+                        state[i*3+0] = torch.zeros(args.n_embd, dtype=atype, requires_grad=False, device=dev).contiguous()
+                        state[i*3+1] = torch.zeros((args.n_head, args.n_att//args.n_head, args.n_att//args.n_head), dtype=torch.float, requires_grad=False, device=dev).contiguous()
+                        state[i*3+2] = torch.zeros(args.n_embd, dtype=atype, requires_grad=False, device=dev).contiguous()
+
+            if embs is None:
+                seq_mode = len(tokens) > 1
+                x = w['emb.weight'][tokens if seq_mode else tokens[0]]
+            else:
+                x = embs
+                seq_mode = True
+
+            for i in range(args.n_layer):
+                bbb = f'blocks.{i}.'
+                att = f'blocks.{i}.att.'
+                ffn = f'blocks.{i}.ffn.'
+                dd = self.strategy[i]
+                dev = dd.device
+                atype = dd.atype
+                wtype = dd.wtype
+                if seq_mode:
+                    cuda_applicable = os.environ["RWKV_CUDA_ON"] == '1' and 'cuda' in str(dev)
+                    if cuda_applicable:
+                        ATT = self.cuda_att_seq
+                    else:
+                        ATT = self.att_seq
+                    if self.version == 5:
+                        ATT = self.att_seq_v5
+                    elif self.version == 5.1:
+                        ATT = self.att_seq_v5_1
+                    elif self.version == 5.2:
+                        ATT = self.att_seq_v5_2
+                        if cuda_applicable:
+                            ATT = self.cuda_att_seq_v5_2
+                    elif self.version == 6.0:
+                        ATT = self.att_seq_v6_0
+                        if cuda_applicable:
+                            ATT = self.cuda_att_seq_v6_0
+                    FFN = self.ffn_seq
+                    if self.version >= 6.0:
+                        FFN = self.ffn_seq_v6
+                else:
+                    ATT = self.att_one
+                    if self.version == 5:
+                        ATT = self.att_one_v5
+                    elif self.version == 5.1:
+                        ATT = self.att_one_v5_1
+                    elif self.version == 5.2:
+                        ATT = self.att_one_v5_1 # same as v5.1
+                    elif self.version == 6.0:
+                        ATT = self.att_one_v6_0
+                    FFN = self.ffn_one
+                    if self.version >= 6.0:
+                        FFN = self.ffn_one_v6
+
+                x = x.to(dtype=atype, device=dev)
+
+                kw = w[f'{att}key.weight']
+                vw = w[f'{att}value.weight']
+                rw = w[f'{att}receptance.weight']
+                ow = w[f'{att}output.weight']
+                if dd.stream:
+                    kw = kw.to(device=dev, non_blocking=True)
+                    vw = vw.to(device=dev, non_blocking=True)
+                    rw = rw.to(device=dev, non_blocking=True)
+                    ow = ow.to(device=dev, non_blocking=True)
+                kmx = w[f'{att}key.weight_mx'] if wtype == torch.uint8 else x
+                krx = w[f'{att}key.weight_rx'] if wtype == torch.uint8 else x
+                kmy = w[f'{att}key.weight_my'] if wtype == torch.uint8 else x
+                kry = w[f'{att}key.weight_ry'] if wtype == torch.uint8 else x
+                vmx = w[f'{att}value.weight_mx'] if wtype == torch.uint8 else x
+                vrx = w[f'{att}value.weight_rx'] if wtype == torch.uint8 else x
+                vmy = w[f'{att}value.weight_my'] if wtype == torch.uint8 else x
+                vry = w[f'{att}value.weight_ry'] if wtype == torch.uint8 else x
+                rmx = w[f'{att}receptance.weight_mx'] if wtype == torch.uint8 else x
+                rrx = w[f'{att}receptance.weight_rx'] if wtype == torch.uint8 else x
+                rmy = w[f'{att}receptance.weight_my'] if wtype == torch.uint8 else x
+                rry = w[f'{att}receptance.weight_ry'] if wtype == torch.uint8 else x
+                omx = w[f'{att}output.weight_mx'] if wtype == torch.uint8 else x
+                orx = w[f'{att}output.weight_rx'] if wtype == torch.uint8 else x
+                omy = w[f'{att}output.weight_my'] if wtype == torch.uint8 else x
+                ory = w[f'{att}output.weight_ry'] if wtype == torch.uint8 else x
+                if self.version in [5.1, 5.2, 6.0]:
+                    gw = w[f'{att}gate.weight']
+                    if dd.stream:
+                        gw = gw.to(device=dev, non_blocking=True)
+                    gmx = w[f'{att}gate.weight_mx'] if wtype == torch.uint8 else x
+                    grx = w[f'{att}gate.weight_rx'] if wtype == torch.uint8 else x
+                    gmy = w[f'{att}gate.weight_my'] if wtype == torch.uint8 else x
+                    gry = w[f'{att}gate.weight_ry'] if wtype == torch.uint8 else x
+                if self.version == 4:
+                    x, state[i*5+0], state[i*5+1], state[i*5+2], state[i*5+3] = ATT(
+                        x, state[i*5+0], state[i*5+1], state[i*5+2], state[i*5+3],
+                        w[f'{bbb}ln1.weight'], w[f'{bbb}ln1.bias'],
+                        w[f'{att}time_mix_k'], w[f'{att}time_mix_v'], w[f'{att}time_mix_r'],
+                        w[f'{att}time_decay'], w[f'{att}time_first'],
+                        kw, vw, rw, ow,
+                        kmx, krx, kmy, kry,
+                        vmx, vrx, vmy, vry,
+                        rmx, rrx, rmy, rry,
+                        omx, orx, omy, ory,
+                        )
+                elif self.version == 5:
+                    x, state[i*3+0], state[i*3+1] = ATT(
+                        x, state[i*3+0], state[i*3+1],
+                        w[f'{bbb}ln1.weight'], w[f'{bbb}ln1.bias'],
+                        w[f'{att}ln_x.weight'], w[f'{att}ln_x.bias'],
+                        w[f'{att}time_mix_k'], w[f'{att}time_mix_v'], w[f'{att}time_mix_r'],
+                        w[f'{att}time_decay'], w[f'{att}time_first'],
+                        kw, vw, rw, ow,
+                        kmx, krx, kmy, kry,
+                        vmx, vrx, vmy, vry,
+                        rmx, rrx, rmy, rry,
+                        omx, orx, omy, ory,
+                        )
+                elif self.version in [5.1, 5.2]:
+                    x, state[i*3+0], state[i*3+1] = ATT(
+                        x, state[i*3+0], state[i*3+1],
+                        w[f'{bbb}ln1.weight'], w[f'{bbb}ln1.bias'],
+                        w[f'{att}ln_x.weight'], w[f'{att}ln_x.bias'],
+                        w[f'{att}time_mix_k'], w[f'{att}time_mix_v'], w[f'{att}time_mix_r'], w[f'{att}time_mix_g'],
+                        w[f'{att}time_decay'], w[f'{att}time_first'],
+                        kw, vw, rw, gw, ow,
+                        kmx, krx, kmy, kry,
+                        vmx, vrx, vmy, vry,
+                        rmx, rrx, rmy, rry,
+                        gmx, grx, gmy, gry,
+                        omx, orx, omy, ory,
+                        )
+                elif self.version == 6.0:
+                    x, state[i*3+0], state[i*3+1] = ATT(
+                        x, state[i*3+0], state[i*3+1],
+                        w[f'{bbb}ln1.weight'], w[f'{bbb}ln1.bias'],
+                        w[f'{att}ln_x.weight'], w[f'{att}ln_x.bias'],
+                        w[f'{att}time_maa_x'], w[f'{att}time_maa_w'], w[f'{att}time_maa_k'], w[f'{att}time_maa_v'], w[f'{att}time_maa_r'], w[f'{att}time_maa_g'],
+                        w[f'{att}time_maa_w1'], w[f'{att}time_maa_w2'], w[f'{att}time_decay_w1'], w[f'{att}time_decay_w2'],
+                        w[f'{att}time_decay'], w[f'{att}time_first'],
+                        kw, vw, rw, gw, ow,
+                        kmx, krx, kmy, kry,
+                        vmx, vrx, vmy, vry,
+                        rmx, rrx, rmy, rry,
+                        gmx, grx, gmy, gry,
+                        omx, orx, omy, ory,
+                        )
+                if dd.stream:
+                    del kw, vw, rw, ow
+                    if self.version in [5.1, 5.2, 6.0]:
+                        del gw
+
+                kw = w[f'{ffn}key.weight']
+                vw = w[f'{ffn}value.weight']
+                rw = w[f'{ffn}receptance.weight']
+                if dd.stream:
+                    kw = kw.to(device=dev, non_blocking=True)
+                    vw = vw.to(device=dev, non_blocking=True)
+                    rw = rw.to(device=dev, non_blocking=True)
+                kmx = w[f'{ffn}key.weight_mx'] if wtype == torch.uint8 else x
+                krx = w[f'{ffn}key.weight_rx'] if wtype == torch.uint8 else x
+                kmy = w[f'{ffn}key.weight_my'] if wtype == torch.uint8 else x
+                kry = w[f'{ffn}key.weight_ry'] if wtype == torch.uint8 else x
+                vmx = w[f'{ffn}value.weight_mx'] if wtype == torch.uint8 else x
+                vrx = w[f'{ffn}value.weight_rx'] if wtype == torch.uint8 else x
+                vmy = w[f'{ffn}value.weight_my'] if wtype == torch.uint8 else x
+                vry = w[f'{ffn}value.weight_ry'] if wtype == torch.uint8 else x
+                rmx = w[f'{ffn}receptance.weight_mx'] if wtype == torch.uint8 else x
+                rrx = w[f'{ffn}receptance.weight_rx'] if wtype == torch.uint8 else x
+                rmy = w[f'{ffn}receptance.weight_my'] if wtype == torch.uint8 else x
+                rry = w[f'{ffn}receptance.weight_ry'] if wtype == torch.uint8 else x
+                if self.version == 4:
+                    offset = i*5+4
+                elif int(self.version) in [5,6]:
+                    offset = i*3+2
+                if self.version < 6.0:
+                    x, state[offset] = FFN(
+                        x, state[offset],
+                        w[f'{bbb}ln2.weight'], w[f'{bbb}ln2.bias'],
+                        w[f'{ffn}time_mix_k'], w[f'{ffn}time_mix_r'],
+                        kw, vw, rw,
+                        kmx, krx, kmy, kry,
+                        vmx, vrx, vmy, vry,
+                        rmx, rrx, rmy, rry,                    
+                        )
+                else:
+                    x, state[offset] = FFN(
+                        x, state[offset],
+                        w[f'{bbb}ln2.weight'], w[f'{bbb}ln2.bias'],
+                        w[f'{ffn}time_maa_k'], w[f'{ffn}time_maa_r'],
+                        kw, vw, rw,
+                        kmx, krx, kmy, kry,
+                        vmx, vrx, vmy, vry,
+                        rmx, rrx, rmy, rry,                    
+                        )
+                if dd.stream:                
+                    del kw, vw, rw
+                
+                if self.RESCALE_LAYER > 0:
+                    if (i+1) % self.RESCALE_LAYER == 0:
+                        x = x / 2
+            
+            dd = self.strategy[args.n_layer]
+            x = x[-1,:] if (seq_mode and (not full_output)) else x
+            x = x.to(dtype=dd.atype, device=dd.device)
+            
+            x = F.layer_norm(x, (args.n_embd,), weight=w['ln_out.weight'], bias=w['ln_out.bias'])
+            if w['head.weight'].dtype != torch.uint8:
+                x = x @ w['head.weight']
+            else:
+                if seq_mode and full_output:
+                    x = mm8_seq(x, w['head.weight'], w['head.weight_mx'], w['head.weight_rx'], w['head.weight_my'], w['head.weight_ry'])
+                else:
+                    x = mm8_one(x, w['head.weight'], w['head.weight_mx'], w['head.weight_rx'], w['head.weight_my'], w['head.weight_ry'])
+
+            return x.float(), state

modeling_vision.py

@@ -0,0 +1,48 @@
+from transformers import CLIPVisionModel
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from dataclasses import dataclass
+
+@dataclass
+class VisionEncoderConfig:
+    n_embd: int = 2048
+    vision_tower_name: str = 'openai/clip-vit-large-patch14-336'
+    grid_size: int = -1 # -1: no grid pooling, 0: take cls token, 1: global avg pooling, 2, 3, 4, ...: grid pooling
+
+class VisionEncoder(nn.Module):
+    def __init__(self, args):
+        super().__init__()
+        self.args = args
+        self.vit = CLIPVisionModel.from_pretrained(args.vision_tower_name)
+        self.proj = nn.Linear(self.vit.config.hidden_size, args.n_embd, bias=False)
+
+    def encode_images(self, images):
+        B, N, C, H, W = images.shape
+        images = images.view(B*N, C, H, W)
+        image_features = self.vit(images).last_hidden_state
+        L, D = image_features.shape[1], image_features.shape[2]
+        # rerange [B*N, L, D] -> [B, N, L, D]
+        image_features = image_features.view(B, N, L, D)[:, 0, :, :]
+        image_features = self.grid_pooling(image_features)
+        return self.proj(image_features)
+    
+    def grid_pooling(self, image_features):
+        if self.args.grid_size == -1: # no grid pooling
+            return image_features
+        if self.args.grid_size == 0: # take cls token
+            return image_features[:, 0:1, :]
+        if self.args.grid_size == 1: # global avg pooling
+            return image_features.mean(dim=1, keepdim=True)
+        cls_features = image_features[:, 0:1, :]
+        image_features = image_features[:, 1:, :] #drop cls token
+        B, L, D = image_features.shape
+        H_or_W = int(L**0.5)
+        image_features = image_features.view(B, H_or_W, H_or_W, D)
+        grid_stride = H_or_W // self.args.grid_size
+        image_features = F.avg_pool2d(image_features.permute(0, 3, 1, 2), 
+                                      padding=0,
+                                      kernel_size=grid_stride, 
+                                      stride=grid_stride)
+        image_features = image_features.permute(0, 2, 3, 1).view(B, -1, D)
+        return torch.cat((cls_features, image_features), dim=1)

requirements.txt

@@ -1,8 +1,7 @@
-gradio==3.28.1
 torch
+transformers
 ninja
 tokenizers
-rwkv==0.8.21
+rwkv==0.8.22
 pynvml
-huggingface_hub
-gradio==3.28.1
+huggingface_hub