release v1.4.7

add translation.

.github/workflows/release.yml

@@ -108,6 +108,7 @@ jobs:
           cd ..
           go install github.com/wailsapp/wails/v2/cmd/wails@latest
           rm -rf ./backend-python/wkv_cuda_utils
+          rm ./backend-python/rwkv_pip/beta/wkv_cuda.pyd
           rm ./backend-python/get-pip.py
           sed -i '1,2d' ./backend-golang/wsl_not_windows.go
           rm ./backend-golang/wsl.go
@@ -140,6 +141,7 @@ jobs:
           cd ..
           go install github.com/wailsapp/wails/v2/cmd/wails@latest
           rm -rf ./backend-python/wkv_cuda_utils
+          rm ./backend-python/rwkv_pip/beta/wkv_cuda.pyd
           rm ./backend-python/get-pip.py
           sed -i '' '1,2d' ./backend-golang/wsl_not_windows.go
           rm ./backend-golang/wsl.go

.gitignore

@@ -26,3 +26,4 @@ __pycache__
 train_log.txt
 finetune/json2binidx_tool/data
 /wsl.state
+/components

CURRENT_CHANGE.md

@@ -1,14 +1,8 @@
 ## Changes
 
-- frontend: update manifest (a lot of new models)
-- frontend: correct Preset UI description
-- frontend: add HardwareMonitor (Windows Only)
-- lora finetune: fix max_epochs (#170)
-- python-backend: allow message content to be empty
-- python-backend: extra ChatCompletionBody params (`raw`, `presystem`)
-- python-backend: add default_stop when stop is null
-- webgpu: fix webgpu_server file permissions of linux and macos
-- chore
+- custom tokenizer .txt support
+- add pre-compiled beta cuda kernel (rwkv-beta==0.8.5, 40%+ faster for fp16) (thanks to #180, pre-compiled kernel of RTX
+  40 Series will be included later)
 
 ## Install
 

README_JA.md

@@ -91,8 +91,8 @@ body.json:
 
 ## 埋め込み API の例
 
-Note: v1.4.0 has improved the quality of embeddings API. The generated results are not compatible
-with previous versions. If you are using embeddings API to generate knowledge bases or similar, please regenerate.
+注意: v1.4.0 では、埋め込み API の品質が向上しました。生成される結果は、以前のバージョンとは互換性がありません。
+もし、embeddings API を使って知識ベースなどを生成している場合は、再生成してください。
 
 LangChain を使用している場合は、`OpenAIEmbeddings(openai_api_base="http://127.0.0.1:8000", openai_api_key="sk-")`
 を使用してください

backend-python/routes/completion.py

@@ -40,7 +40,7 @@ default_stop = [
 
 class ChatCompletionBody(ModelConfigBody):
     messages: Union[List[Message], None]
-    model: str = "rwkv"
+    model: Union[str, None] = "rwkv"
     stream: bool = False
     stop: Union[str, List[str], None] = default_stop
     user_name: Union[str, None] = Field(None, description="Internal user name")
@@ -74,7 +74,7 @@ class ChatCompletionBody(ModelConfigBody):
 
 class CompletionBody(ModelConfigBody):
     prompt: Union[str, List[str], None]
-    model: str = "rwkv"
+    model: Union[str, None] = "rwkv"
     stream: bool = False
     stop: Union[str, List[str], None] = None
 
@@ -368,7 +368,7 @@ async def completions(body: CompletionBody, request: Request):
 
 class EmbeddingsBody(BaseModel):
     input: Union[str, List[str], List[List[int]], None]
-    model: str = "rwkv"
+    model: Union[str, None] = "rwkv"
     encoding_format: str = None
     fast_mode: bool = False
 

backend-python/routes/config.py

@@ -29,6 +29,7 @@ def get_tokens_path(model_path: str):
 class SwitchModelBody(BaseModel):
     model: str
     strategy: str
+    tokenizer: Union[str, None] = None
     customCuda: bool = False
 
     class Config:
@@ -36,6 +37,7 @@ class SwitchModelBody(BaseModel):
             "example": {
                 "model": "models/RWKV-4-World-3B-v1-20230619-ctx4096.pth",
                 "strategy": "cuda fp16",
+                "tokenizer": None,
                 "customCuda": False,
             }
         }
@@ -65,19 +67,24 @@ def switch_model(body: SwitchModelBody, response: Response, request: Request):
     os.environ["RWKV_CUDA_ON"] = "1" if body.customCuda else "0"
 
     global_var.set(global_var.Model_Status, global_var.ModelStatus.Loading)
+    tokenizer = (
+        get_tokens_path(body.model)
+        if body.tokenizer is None or body.tokenizer == ""
+        else body.tokenizer
+    )
     try:
         global_var.set(
             global_var.Model,
             TextRWKV(
                 model=body.model,
                 strategy=body.strategy,
-                tokens_path=get_tokens_path(body.model),
+                tokens_path=tokenizer,
             )
             if "midi" not in body.model.lower()
             else MusicRWKV(
                 model=body.model,
                 strategy=body.strategy,
-                tokens_path=get_tokens_path(body.model),
+                tokens_path=tokenizer,
             ),
         )
     except Exception as e:

backend-python/rwkv_pip/beta/cuda/att_one.cu

@@ -88,7 +88,7 @@ struct Mix {
 
 using torch::Tensor;
 
-void gemm_fp16_cublas(Tensor a, Tensor b, Tensor c);
+void gemm_fp16_cublas_tensor(Tensor a, Tensor b, Tensor c);
 
 Tensor att_one(Tensor x, Tensor ln_w, Tensor ln_b, Tensor sx, Tensor k_mix,
              Tensor v_mix, Tensor r_mix, Tensor kw,
@@ -105,9 +105,9 @@ Tensor att_one(Tensor x, Tensor ln_w, Tensor ln_b, Tensor sx, Tensor k_mix,
                    data_ptr<half>(vx), data_ptr<half>(rx)},
                x.numel());
 
-  gemm_fp16_cublas(kx, kw, k);
-  gemm_fp16_cublas(vx, vw, v);
-  gemm_fp16_cublas(rx, rw, r);
+  gemm_fp16_cublas_tensor(kx, kw, k);
+  gemm_fp16_cublas_tensor(vx, vw, v);
+  gemm_fp16_cublas_tensor(rx, rw, r);
   at::sigmoid_(r);
 
   element_wise(WkvForwardOne{data_ptr<float>(t_first), data_ptr<float>(k),
@@ -118,7 +118,7 @@ Tensor att_one(Tensor x, Tensor ln_w, Tensor ln_b, Tensor sx, Tensor k_mix,
                              data_ptr<half>(r)},
                x.numel());
 
-  gemm_fp16_cublas(r, ow, x_plus_out);
+  gemm_fp16_cublas_tensor(r, ow, x_plus_out);
   x_plus_out += x;
   return xx;
 }

backend-python/rwkv_pip/beta/cuda/att_one_v5.cu

@@ -0,0 +1,109 @@
+#include "ATen/ATen.h"
+#include <cuda_fp16.h>
+#include <cuda_runtime.h>
+#include <torch/extension.h>
+
+#include "element_wise.h"
+#include "util.h"
+
+// Equivalent Python code:
+// s1 = t_first * a + s
+// s2 = a + t_decay * s
+struct Fused1 {
+  const float *t_first;
+  const float *t_decay;
+  const float *a;
+  const float *s;
+  const int32_t inner_size;
+  /* out */ float *s1;
+  /* out */ float *s2;
+
+  __device__ void operator()(int i) const {
+    const int j = i / inner_size;
+    s1[i] = t_first[j] * a[i] + s[i];
+    s2[i] = a[i] + t_decay[j] * s[i];
+  }
+};
+
+/*
+   Equivalent Python code:
+   kx = xx * k_mix + sx * (1 - k_mix)
+   vx = xx * v_mix + sx * (1 - v_mix)
+   rx = xx * r_mix + sx * (1 - r_mix)
+*/
+
+struct Mix {
+  const half *xx;
+  const half *sx;
+  const half *k_mix;
+  const half *v_mix;
+  const half *r_mix;
+  /* out */ half *kx;
+  /* out */ half *vx;
+  /* out */ half *rx;
+
+  __device__ void operator()(int i) const {
+    half xx_ = xx[i];
+    half sx_ = sx[i];
+    half k_mix_ = k_mix[i];
+    half v_mix_ = v_mix[i];
+    half r_mix_ = r_mix[i];
+    kx[i] = __hadd(__hmul(xx_, k_mix_),
+                   __hmul(sx_, __hsub(__float2half(1), k_mix_)));
+    vx[i] = __hadd(__hmul(xx_, v_mix_),
+                   __hmul(sx_, __hsub(__float2half(1), v_mix_)));
+    rx[i] = __hadd(__hmul(xx_, r_mix_),
+                   __hmul(sx_, __hsub(__float2half(1), r_mix_)));
+  }
+};
+
+using torch::Tensor;
+
+void gemm_fp16_cublas_tensor(Tensor a, Tensor b, Tensor c);
+
+Tensor att_one_v5(Tensor x, Tensor sx, Tensor s, Tensor ln_w, Tensor ln_b,
+                  Tensor lx_w, Tensor lx_b, Tensor k_mix, Tensor v_mix,
+                  Tensor r_mix, Tensor kw,
+                  /* imm */ Tensor kx, Tensor vw, /* imm */ Tensor vx,
+                  Tensor rw,
+                  /* imm */ Tensor rx, Tensor ow, Tensor t_first,
+                  /* imm */ Tensor k, Tensor t_decay, /* imm */ Tensor v,
+                  /* imm */ Tensor r, /* imm */ Tensor s1,
+                  /* out */ Tensor x_plus_out, /* out */ Tensor s2) {
+  Tensor xx = at::layer_norm(x, {x.size(-1)}, ln_w, ln_b);
+  element_wise(Mix{data_ptr<half>(xx), data_ptr<half>(sx),
+                   data_ptr<half>(k_mix), data_ptr<half>(v_mix),
+                   data_ptr<half>(r_mix), data_ptr<half>(kx),
+                   data_ptr<half>(vx), data_ptr<half>(rx)},
+               x.numel());
+
+  int H = t_decay.size(0);
+  int S = x.size(-1) / H;
+  gemm_fp16_cublas_tensor(rx, rw, r);
+  r = at::reshape(r, {H, 1, S});
+  gemm_fp16_cublas_tensor(kx, kw, k);
+  k = at::reshape(k, {H, S, 1});
+  gemm_fp16_cublas_tensor(vx, vw, v);
+  v = at::reshape(v, {H, 1, S});
+
+  {
+    Tensor a = at::matmul(k, v);
+
+    // s1 = t_first * a + s
+    // s2 = a + t_decay * s
+    element_wise(Fused1{data_ptr<float>(t_first), data_ptr<float>(t_decay),
+                        data_ptr<float>(a), data_ptr<float>(s),
+                        static_cast<int32_t>(a.size(1) * a.size(2)),
+                        data_ptr<float>(s1), data_ptr<float>(s2)},
+                 a.numel());
+  }
+
+  Tensor out = at::matmul(r, s1);
+  out = at::flatten(out);
+  out = at::squeeze(at::group_norm(at::unsqueeze(out, 0), H, lx_w, lx_b), 0);
+  out = at::_cast_Half(out);
+
+  gemm_fp16_cublas_tensor(out, ow, x_plus_out);
+  x_plus_out += x;
+  return xx;
+}

backend-python/rwkv_pip/beta/cuda/att_seq.cu

@@ -8,7 +8,6 @@
 
 using torch::Tensor;
 
-void gemm_fp16_cublas(Tensor a, Tensor b, Tensor c);
 void gemm_fp16_cublas(const void *a, const void *b, void *c, int m,
                       int n, int k, bool output_fp32);
 

backend-python/rwkv_pip/beta/cuda/gemm_fp16_cublas.cpp

@@ -70,11 +70,59 @@ void gemm_fp16_cublas(const void *a, const void *b, void *c, int ori_m,
       cuda_c_data_type, cublas_ldc, compute_type, algo));
 }
 
-void gemm_fp16_cublas(torch::Tensor a, torch::Tensor b, torch::Tensor c) {
-  // comptiable with rwkv one mode, 1-D tensor * 2-D tensor
-  const int m = a.dense_dim() == 1 ? 1 : a.size(0);
-  const int n = b.size(1);
-  const int k = b.size(0);
-  gemm_fp16_cublas(a.data_ptr(), b.data_ptr(), c.data_ptr(), m, n, k,
-                   c.dtype() == torch::kFloat32);
+/*
+  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_tensor(torch::Tensor a, torch::Tensor b, torch::Tensor c) {
+  if (a.sizes().size() == 1) {
+    assert(b.sizes().size() == 2);
+    a = at::unsqueeze(a, 0);
+  }
+  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 = get_cublas_handle();
+
+#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));
+  }
 }

backend-python/rwkv_pip/beta/cuda/wrapper.cpp

@@ -118,7 +118,9 @@ void mm8_one(int64_t N, int64_t M,
 
 using torch::Tensor;
 
-void gemm_fp16_cublas(Tensor a, Tensor b, Tensor c);
+#ifndef DISABLE_CUBLAS_GEMM
+void gemm_fp16_cublas_tensor(Tensor a, Tensor b, Tensor c);
+#endif
 
 Tensor att_one(Tensor x, Tensor ln_w, Tensor ln_b, Tensor sx, Tensor k_mix,
              Tensor v_mix, Tensor r_mix, Tensor kw,
@@ -134,6 +136,16 @@ Tensor att_seq(Tensor x, Tensor sx, Tensor ln_w, Tensor ln_b, Tensor k_mix,
                Tensor ow, Tensor t_first, Tensor pp, Tensor aa, Tensor bb,
                Tensor t_decay, /* imm */ Tensor buf, /* out */ Tensor x_plus_out);
 
+Tensor att_one_v5(Tensor x, Tensor sx, Tensor s, Tensor ln_w, Tensor ln_b,
+                  Tensor lx_w, Tensor lx_b, Tensor k_mix, Tensor v_mix,
+                  Tensor r_mix, Tensor kw,
+                  /* imm */ Tensor kx, Tensor vw, /* imm */ Tensor vx,
+                  Tensor rw,
+                  /* imm */ Tensor rx, Tensor ow, Tensor t_first,
+                  /* imm */ Tensor k, Tensor t_decay, /* imm */ Tensor v,
+                  /* imm */ Tensor r, /* imm */ Tensor s1,
+                  /* out */ Tensor x_plus_out, /* out */ Tensor s2);
+
 Tensor ffn_seq(Tensor x, Tensor sx, Tensor ln_w, Tensor ln_b, Tensor k_mix,
                Tensor r_mix, Tensor kw, Tensor vw, Tensor rw,
                /* imm */ Tensor buf,
@@ -148,8 +160,9 @@ 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");
-    m.def("gemm_fp16_cublas", &gemm_fp16_cublas, "gemv fp16 cublas");
+    m.def("gemm_fp16_cublas", &gemm_fp16_cublas_tensor, "gemv fp16 cublas");
     m.def("att_one", &att_one, "att one");
+    m.def("att_one_v5", &att_one_v5, "att one v5");
     m.def("att_seq", &att_seq, "att seq");
     m.def("ffn_seq", &ffn_seq, "ffn seq");
     m.def("ffn_one", &ffn_one, "ffn one");
@@ -159,8 +172,9 @@ TORCH_LIBRARY(rwkv, m) {
     m.def("wkv_forward", wkv_forward);
     m.def("mm8_seq", mm8_seq);
     m.def("mm8_one", mm8_one);
-    m.def("gemm_fp16_cublas", gemm_fp16_cublas);
+    m.def("gemm_fp16_cublas", gemm_fp16_cublas_tensor);
     m.def("att_one", att_one);
+    m.def("att_one_v5", &att_one_v5);
     m.def("att_seq", att_seq);
     m.def("ffn_seq", ffn_seq);
     m.def("ffn_one", ffn_one);

backend-python/rwkv_pip/beta/model.py

@@ -3,7 +3,7 @@
 ########################################################################################################
 
 from typing import Optional
-import types, gc, os, time, re
+import types, gc, os, time, re, platform
 import torch
 from torch.nn import functional as F
 
@@ -91,8 +91,10 @@ if os.environ.get("RWKV_CUDA_ON") == "1":
                 f"{current_path}/cuda/att_one.cu",
                 f"{current_path}/cuda/att_seq.cu",
                 f"{current_path}/cuda/ffn.cu",
+                f"{current_path}/cuda/att_one_v5.cu",
             ],
             verbose=True,
+            extra_ldflags=["cublas.lib"],
             extra_cuda_cflags=[
                 "-t 4",
                 "-std=c++17",
@@ -149,26 +151,40 @@ if os.environ.get("RWKV_CUDA_ON") == "1":
         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"
+
+if os.environ.get("RWKV_CUDA_ON") == "1":
+
     @MyStatic
     def gemm(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":
-            assert len(b.shape) == 2
             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:
-    os.environ["RWKV_CUDA_ON"] = "0"
 
     def gemm(a, b, output_dtype: Optional[torch.dtype] = None):
         if output_dtype is None:
@@ -217,7 +233,7 @@ class RWKV(MyModule):
             )  # load model to CPU first
             # it is supported to load a pure meta-tensor state dict (e.g. for quick testing)
             for k, v in self.w.items():
-                if v.is_meta:
+                if isinstance(v, torch.Tensor) and v.is_meta:
                     # torch.zeros_like(v, device='cpu') doesn't produce an all-zero tensor
                     # if v is a meta tensor
                     self.w[k] = torch.zeros(v.shape, dtype=v.dtype, device="cpu")
@@ -247,9 +263,14 @@ class RWKV(MyModule):
             args.n_embd = w["emb.weight"].shape[1]
             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 = 5
+                if self.version == 5 and "att.time_decay" in x:
+                    args.n_head = w[x].shape[0]
 
             ####################### Compute strategy
 
@@ -352,6 +373,20 @@ class RWKV(MyModule):
                 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
@@ -382,8 +417,19 @@ class RWKV(MyModule):
                         w[x] = w[x].t()
 
                     if ".time_decay" in x:  # need fp32 for this
+                        if self.version == 4:
                             w[x] = -torch.exp(w[x].float())
+                        elif self.version == 5:
+                            w[x] = torch.exp(-torch.exp(w[x].float())).reshape(-1, 1, 1)
                     elif ".time_first" in x:  # need fp32 for this
+                        if self.version == 4:
+                            w[x] = w[x].float()
+                        elif self.version == 5:
+                            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)
+                    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):
@@ -931,6 +977,147 @@ class RWKV(MyModule):
 
     ########################################################################################################
 
+    @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]
+        S = x.shape[-1] // H
+
+        r = gemm(rx, rw, output_dtype=torch.float32).view(H, 1, S)
+        k = gemm(kx, kw, output_dtype=torch.float32).view(H, S, 1)
+        v = gemm(vx, vw, output_dtype=torch.float32).view(H, 1, S)
+
+        a = gemm(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
+        ).squeeze(0)
+        out = out.to(dtype=x.dtype)
+        out = gemm(out, ow)
+
+        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]
+        S = 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 = gemm(rx, rw, output_dtype=torch.float32).view(T, H, S).transpose(0, 1)
+        k = (
+            gemm(kx, kw, output_dtype=torch.float32)
+            .view(T, H, S)
+            .transpose(0, 1)
+            .transpose(-2, -1)
+        )
+        v = gemm(vx, vw, output_dtype=torch.float32).view(T, H, S).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 * S)
+        out = F.group_norm(out, num_groups=H, weight=lx_w, bias=lx_b)
+        out = out.to(dtype=x.dtype)
+        out = gemm(out, ow)
+
+        return x + out, xx[-1, :], s
+
+    ########################################################################################################
+
     if os.environ["RWKV_CUDA_ON"] == "1":
 
         @MyFunction
@@ -1140,7 +1327,7 @@ class RWKV(MyModule):
             xx = torch.ops.rwkv.ffn_seq(
                 x, sx, ln_w, ln_b, k_mix, r_mix, kw, vw, rw, buf, x_plus_out
             )
-            return x_plus_out, xx[-1:]
+            return x_plus_out, xx[-1, :]
 
         @MyFunction
         def cuda_att_one_fp16(
@@ -1220,6 +1407,86 @@ class RWKV(MyModule):
             )
             return x_plus_out_t, xx, t1_t, t2_t, p_t
 
+        @MyFunction
+        def cuda_att_one_v5_fp16(
+            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,
+        ):
+            kx = torch.empty_like(x)
+            vx = torch.empty_like(x)
+            rx = torch.empty_like(x)
+
+            H = t_decay.shape[0]
+            S = x.shape[-1] // H
+
+            r = torch.empty((H * S,), dtype=torch.float32, device=x.device)
+            k = torch.empty((H * S,), dtype=torch.float32, device=x.device)
+            v = torch.empty((H * S,), dtype=torch.float32, device=x.device)
+            s1 = torch.empty((H, S, S), dtype=torch.float32, device=x.device)
+            s2 = torch.empty((H, S, S), dtype=torch.float32, device=x.device)
+            x_plus_out = torch.empty_like(x)
+
+            xx = torch.ops.rwkv.att_one_v5(
+                x,
+                sx,
+                s,
+                ln_w,
+                ln_b,
+                lx_w,
+                lx_b,
+                k_mix,
+                v_mix,
+                r_mix,
+                kw,
+                kx,
+                vw,
+                vx,
+                rw,
+                rx,
+                ow,
+                t_first,
+                k,
+                t_decay,
+                v,
+                r,
+                s1,
+                x_plus_out,
+                s2,
+            )
+
+            return x_plus_out, xx, s2
+
         @MyFunction
         def cuda_ffn_one_fp16(
             self,
@@ -1265,6 +1532,7 @@ class RWKV(MyModule):
             args = self.args
 
             if state == None:
+                if self.version == 4:
                     state = [None] * args.n_layer * 5
                     for i in range(
                         args.n_layer
@@ -1276,10 +1544,16 @@ class RWKV(MyModule):
                             args.n_embd, dtype=atype, requires_grad=False, device=dev
                         ).contiguous()
                         state[i * 5 + 1] = torch.zeros(
-                        args.n_embd, dtype=torch.float, requires_grad=False, device=dev
+                            args.n_embd,
+                            dtype=torch.float,
+                            requires_grad=False,
+                            device=dev,
                         ).contiguous()
                         state[i * 5 + 2] = torch.zeros(
-                        args.n_embd, dtype=torch.float, requires_grad=False, device=dev
+                            args.n_embd,
+                            dtype=torch.float,
+                            requires_grad=False,
+                            device=dev,
                         ).contiguous()
                         state[i * 5 + 3] = (
                             torch.zeros(
@@ -1293,6 +1567,28 @@ class RWKV(MyModule):
                         state[i * 5 + 4] = torch.zeros(
                             args.n_embd, dtype=atype, requires_grad=False, device=dev
                         ).contiguous()
+                elif self.version == 5:
+                    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_embd // args.n_head,
+                                args.n_embd // 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()
 
             seq_mode = len(tokens) > 1
 
@@ -1317,9 +1613,13 @@ class RWKV(MyModule):
                             ATT = self.cuda_att_seq_i8
                         else:
                             ATT = self.cuda_att_seq_naive
+                    if self.version == 5:
+                        ATT = self.att_seq_v5
                 else:
                     ATT = self.att_one if wtype != torch.uint8 else self.att_one_i8
                     FFN = self.ffn_one if wtype != torch.uint8 else self.ffn_one_i8
+                    if self.version == 5:
+                        ATT = self.att_one_v5
                     if (
                         "cuda" in str(dev)
                         and os.environ["RWKV_CUDA_ON"] == "1"
@@ -1327,6 +1627,8 @@ class RWKV(MyModule):
                     ):
                         ATT = self.cuda_att_one_fp16
                         FFN = self.cuda_ffn_one_fp16
+                        if self.version == 5:
+                            ATT = self.cuda_att_one_v5_fp16
 
                 x = x.to(dtype=atype, device=dev)
 
@@ -1355,6 +1657,7 @@ class RWKV(MyModule):
                 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 == 4:
                     (
                         x,
                         state[i * 5 + 0],
@@ -1395,6 +1698,41 @@ class RWKV(MyModule):
                         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,
+                    )
                 if dd.stream:
                     del kw, vw, rw, ow
 
@@ -1417,9 +1755,13 @@ class RWKV(MyModule):
                 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
-                x, state[i * 5 + 4] = FFN(
+                if self.version == 4:
+                    offset = i * 5 + 4
+                elif self.version == 5:
+                    offset = i * 3 + 2
+                x, state[offset] = FFN(
                     x,
-                    state[i * 5 + 4],
+                    state[offset],
                     w[f"{bbb}ln2.weight"],
                     w[f"{bbb}ln2.bias"],
                     w[f"{ffn}time_mix_k"],

backend-python/rwkv_pip/utils.py

@@ -33,7 +33,7 @@ class PIPELINE_ARGS:
 
 
 class PIPELINE:
-    def __init__(self, model, WORD_NAME):
+    def __init__(self, model, WORD_NAME: str):
         self.model = model
         if WORD_NAME == "cl100k_base":
             import tiktoken
@@ -46,6 +46,12 @@ class PIPELINE:
             self.tokenizer = TRIE_TOKENIZER(
                 os.path.dirname(os.path.abspath(__file__)) + "/rwkv_vocab_v20230424.txt"
             )
+        else:
+            if WORD_NAME.endswith(".txt"):
+                sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+                from rwkv_tokenizer import TRIE_TOKENIZER
+
+                self.tokenizer = TRIE_TOKENIZER(WORD_NAME)
             else:
                 from tokenizers import Tokenizer
 

frontend/src/_locales/ja/main.json

@@ -226,7 +226,7 @@
   "Please select a LoRA model": "LoRAモデルを選択してください",
   "You are using sample data for training. For formal training, please make sure to create your own jsonl file.": "トレーニングにはサンプルデータを使用しています。正式なトレーニングのためには、自身でjsonlファイルを作成してください。",
   "WSL is not running, please retry. If it keeps happening, it means you may be using an outdated version of WSL, run \"wsl --update\" to update.": "WSLが実行されていません、もう一度試してください。これが続く場合、古いバージョンのWSLを使用している可能性があります。\"wsl --update\"を実行して更新してください。",
-  "Memory is not enough, try to increase the virtual memory or use a smaller base model.": "メモリが不足しています、仮想メモリを増やすか小さなベースモデルを使用してみてください。",
+  "Memory is not enough, try to increase the virtual memory (Swap of WSL) or use a smaller base model.": "メモリが不足しています、仮想メモリ (WSL Swap) を増やすか小さなベースモデルを使用してみてください。",
   "VRAM is not enough": "ビデオRAMが不足しています",
   "Training data is not enough, reduce context length or add more data for training": "トレーニングデータが不足しています、コンテキストの長さを減らすか、トレーニング用のデータをさらに追加してください",
   "You are using WSL 1 for training, please upgrade to WSL 2. e.g. Run \"wsl --set-version Ubuntu-22.04 2\"": "トレーニングにWSL 1を使用しています、WSL 2にアップグレードしてください。例:\"wsl --set-version Ubuntu-22.04 2\"を実行する",
@@ -248,5 +248,10 @@
   "Preview Only": "プレビューのみ",
   "RAM": "RAM",
   "VRAM": "VRAM",
-  "GPU Usage": "GPU使用率"
+  "GPU Usage": "GPU使用率",
+  "Use Custom Tokenizer": "カスタムトークナイザーを使用する",
+  "Tokenizer Path (e.g. backend-python/rwkv_pip/20B_tokenizer.json)": "トークナイザーパス (例: backend-python/rwkv_pip/20B_tokenizer.json)",
+  "User Name": "ユーザー名",
+  "Assistant Name": "アシスタント名",
+  "Insert default system prompt at the beginning": "最初にデフォルトのシステムプロンプトを挿入"
 }

frontend/src/_locales/zh-hans/main.json

@@ -226,7 +226,7 @@
   "Please select a LoRA model": "请选择一个LoRA模型",
   "You are using sample data for training. For formal training, please make sure to create your own jsonl file.": "你正在使用示例数据训练，对于正式训练场合，请务必创建你自己的jsonl训练数据",
   "WSL is not running, please retry. If it keeps happening, it means you may be using an outdated version of WSL, run \"wsl --update\" to update.": "WSL没有运行，请重试。如果一直出现此错误，意味着你可能正在使用旧版本的WSL，请在cmd执行\"wsl --update\"以更新",
-  "Memory is not enough, try to increase the virtual memory or use a smaller base model.": "内存不足，尝试增加虚拟内存，或使用一个更小规模的基底模型",
+  "Memory is not enough, try to increase the virtual memory (Swap of WSL) or use a smaller base model.": "内存不足，尝试增加虚拟内存(WSL Swap)，或使用一个更小规模的基底模型",
   "VRAM is not enough": "显存不足",
   "Training data is not enough, reduce context length or add more data for training": "训练数据不足，请减小上下文长度或增加训练数据",
   "You are using WSL 1 for training, please upgrade to WSL 2. e.g. Run \"wsl --set-version Ubuntu-22.04 2\"": "你正在使用WSL 1进行训练，请升级到WSL 2。例如，运行\"wsl --set-version Ubuntu-22.04 2\"",
@@ -248,5 +248,10 @@
   "Preview Only": "仅预览",
   "RAM": "内存",
   "VRAM": "显存",
-  "GPU Usage": "GPU占用"
+  "GPU Usage": "GPU占用",
+  "Use Custom Tokenizer": "使用自定义Tokenizer",
+  "Tokenizer Path (e.g. backend-python/rwkv_pip/20B_tokenizer.json)": "Tokenizer路径 (例如: backend-python/rwkv_pip/20B_tokenizer.json)",
+  "User Name": "用户名称",
+  "Assistant Name": "AI名称",
+  "Insert default system prompt at the beginning": "在开头自动插入默认系统提示"
 }

frontend/src/components/RunButton.tsx

@@ -186,6 +186,7 @@ export const RunButton: FC<{ onClickRun?: MouseEventHandler, iconMode?: boolean
             switchModel({
               model: modelPath,
               strategy: strategy,
+              tokenizer: modelConfig.modelParameters.useCustomTokenizer ? modelConfig.modelParameters.customTokenizer : undefined,
               customCuda: customCudaFile !== ''
             }).then(async (r) => {
               if (r.ok) {

frontend/src/pages/Chat.tsx

@@ -312,7 +312,10 @@ const ChatPanel: FC = observer(() => {
           stream: true,
           model: commonStore.settings.apiChatModelName, // 'gpt-3.5-turbo'
           temperature: apiParams.temperature,
-          top_p: apiParams.topP
+          top_p: apiParams.topP,
+          user_name: commonStore.activePreset?.userName,
+          assistant_name: commonStore.activePreset?.assistantName,
+          presystem: commonStore.activePreset?.presystem
         }),
         signal: chatSseController?.signal,
         onmessage(e) {

frontend/src/pages/Configs.tsx

@@ -1,6 +1,19 @@
-import { Dropdown, Input, Label, Option, Select, Switch, Text } from '@fluentui/react-components';
+import {
+  Accordion,
+  AccordionHeader,
+  AccordionItem,
+  AccordionPanel,
+  Checkbox,
+  Dropdown,
+  Input,
+  Label,
+  Option,
+  Select,
+  Switch,
+  Text
+} from '@fluentui/react-components';
 import { AddCircle20Regular, DataUsageSettings20Regular, Delete20Regular, Save20Regular } from '@fluentui/react-icons';
-import React, { FC } from 'react';
+import React, { FC, useEffect, useRef } from 'react';
 import { Section } from '../components/Section';
 import { Labeled } from '../components/Labeled';
 import { ToolTipButton } from '../components/ToolTipButton';
@@ -43,6 +56,8 @@ export type ModelParameters = {
   maxStoredLayers: number;
   useCustomCuda?: boolean;
   customStrategy?: string;
+  useCustomTokenizer?: boolean;
+  customTokenizer?: string;
 }
 
 export type ModelConfig = {
@@ -57,10 +72,16 @@ export const Configs: FC = observer(() => {
   const [selectedIndex, setSelectedIndex] = React.useState(commonStore.currentModelConfigIndex);
   const [selectedConfig, setSelectedConfig] = React.useState(commonStore.modelConfigs[selectedIndex]);
   const [displayStrategyImg, setDisplayStrategyImg] = React.useState(false);
+  const advancedHeaderRef = useRef<HTMLDivElement>(null);
   const mq = useMediaQuery('(min-width: 640px)');
   const navigate = useNavigate();
   const port = selectedConfig.apiParameters.apiPort;
 
+  useEffect(() => {
+    if (advancedHeaderRef.current)
+      (advancedHeaderRef.current.firstElementChild as HTMLElement).style.padding = '0';
+  }, []);
+
   const updateSelectedIndex = (newIndex: number) => {
     setSelectedIndex(newIndex);
     setSelectedConfig(commonStore.modelConfigs[newIndex]);
@@ -412,6 +433,40 @@ export const Configs: FC = observer(() => {
                         }} />
                     } />
                 }
+                {selectedConfig.modelParameters.device !== 'WebGPU' &&
+                  <Accordion className="sm:col-span-2" collapsible
+                    openItems={!commonStore.modelParamsCollapsed && 'advanced'}
+                    onToggle={(e, data) => {
+                      if (data.value === 'advanced')
+                        commonStore.setModelParamsCollapsed(!commonStore.modelParamsCollapsed);
+                    }}>
+                    <AccordionItem value="advanced">
+                      <AccordionHeader ref={advancedHeaderRef} size="small">{t('Advanced')}</AccordionHeader>
+                      <AccordionPanel>
+                        <div className="flex flex-col">
+                          <div className="flex grow">
+                            <Checkbox className="select-none"
+                              size="large" label={t('Use Custom Tokenizer')}
+                              checked={selectedConfig.modelParameters.useCustomTokenizer}
+                              onChange={(_, data) => {
+                                setSelectedConfigModelParams({
+                                  useCustomTokenizer: data.checked as boolean
+                                });
+                              }} />
+                            <Input className="grow"
+                              placeholder={t('Tokenizer Path (e.g. backend-python/rwkv_pip/20B_tokenizer.json)')!}
+                              value={selectedConfig.modelParameters.customTokenizer}
+                              onChange={(e, data) => {
+                                setSelectedConfigModelParams({
+                                  customTokenizer: data.value
+                                });
+                              }} />
+                          </div>
+                        </div>
+                      </AccordionPanel>
+                    </AccordionItem>
+                  </Accordion>
+                }
               </div>
             }
           />

frontend/src/pages/Downloads.tsx

@@ -1,10 +1,10 @@
-import React, { FC } from 'react';
+import React, { FC, useEffect } from 'react';
 import { useTranslation } from 'react-i18next';
 import { Page } from '../components/Page';
 import { observer } from 'mobx-react-lite';
 import commonStore from '../stores/commonStore';
 import { Divider, Field, ProgressBar } from '@fluentui/react-components';
-import { bytesToGb, bytesToKb, bytesToMb } from '../utils';
+import { bytesToGb, bytesToKb, bytesToMb, refreshLocalModels } from '../utils';
 import { ToolTipButton } from '../components/ToolTipButton';
 import { Folder20Regular, Pause20Regular, Play20Regular } from '@fluentui/react-icons';
 import { AddToDownloadList, OpenFileFolder, PauseDownload } from '../../wailsjs/go/backend_golang/App';
@@ -23,6 +23,12 @@ export type DownloadStatus = {
 
 export const Downloads: FC = observer(() => {
   const { t } = useTranslation();
+  const finishedModelsLen = commonStore.downloadList.filter((status) => status.done && status.name.endsWith('.pth')).length;
+  useEffect(() => {
+    if (finishedModelsLen > 0)
+      refreshLocalModels({ models: commonStore.modelSourceList }, false);
+    console.log('finishedModelsLen:', finishedModelsLen);
+  }, [finishedModelsLen]);
 
   let displayList = commonStore.downloadList.slice();
   const downloadListNames = displayList.map(s => s.name);

frontend/src/pages/PresetsManager/PresetsButton.tsx

@@ -56,6 +56,9 @@ export type Preset = {
   stop: string,
   injectStart: string,
   injectEnd: string,
+  presystem?: boolean,
+  userName?: string,
+  assistantName?: string
 }
 
 export const defaultPreset: Preset = {
@@ -255,9 +258,36 @@ export const ChatPresetEditor: FC<{
             } />
           {
             editingMessages ?
-              <MessagesEditor /> :
+              <div className="flex flex-col gap-1">
+                <Labeled flex spaceBetween label={t('Insert default system prompt at the beginning')}
+                  content={
+                    <Switch checked={editingPreset.presystem === undefined ? true : editingPreset.presystem}
+                      onChange={(e, data) => {
+                        setEditingPreset({
+                          presystem: data.checked
+                        });
+                      }} />
+                  } />
+                <Labeled flex breakline label={t('User Name')}
+                  content={
+                    <Input placeholder="User" value={editingPreset.userName} onChange={(e, data) => {
+                      setEditingPreset({
+                        userName: data.value
+                      });
+                    }} />
+                  } />
+                <Labeled flex breakline label={t('Assistant Name')}
+                  content={
+                    <Input placeholder="Assistant" value={editingPreset.assistantName} onChange={(e, data) => {
+                      setEditingPreset({
+                        assistantName: data.value
+                      });
+                    }} />
+                  } />
+                <MessagesEditor />
+              </div> :
               <div className="flex flex-col gap-1 p-2 overflow-x-hidden overflow-y-auto">
-                <Labeled flex breakline label={`${t('Description')} (${t("Preview Only")})`}
+                <Labeled flex breakline label={`${t('Description')} (${t('Preview Only')})`}
                   content={
                     <Input value={editingPreset.desc} onChange={(e, data) => {
                       setEditingPreset({

frontend/src/pages/Train.tsx

@@ -154,7 +154,7 @@ const showError = (e: any) => {
 };
 
 const errorsMap = Object.entries({
-  'python3 ./finetune/lora/train.py': 'Memory is not enough, try to increase the virtual memory or use a smaller base model.',
+  'python3 ./finetune/lora/train.py': 'Memory is not enough, try to increase the virtual memory (Swap of WSL) or use a smaller base model.',
   'cuda out of memory': 'VRAM is not enough',
   'valueerror: high <= 0': 'Training data is not enough, reduce context length or add more data for training',
   '+= \'+ptx\'': 'You are using WSL 1 for training, please upgrade to WSL 2. e.g. Run "wsl --set-version Ubuntu-22.04 2"',

frontend/src/stores/commonStore.ts

@@ -74,6 +74,7 @@ class CommonStore {
   // configs
   currentModelConfigIndex: number = 0;
   modelConfigs: ModelConfig[] = [];
+  modelParamsCollapsed: boolean = true;
   // models
   modelSourceManifestList: string = 'https://cdn.jsdelivr.net/gh/josstorer/RWKV-Runner@master/manifest.json;';
   modelSourceList: ModelSourceItem[] = [];
@@ -259,6 +260,10 @@ class CommonStore {
     this.advancedCollapsed = value;
   }
 
+  setModelParamsCollapsed(value: boolean) {
+    this.modelParamsCollapsed = value;
+  }
+
   setLastUnfinishedModelDownloads(value: DownloadStatus[]) {
     this.lastUnfinishedModelDownloads = value;
   }

go.mod

@@ -4,15 +4,16 @@ go 1.20
 
 require (
 	github.com/cavaliergopher/grab/v3 v3.0.1
+	github.com/fsnotify/fsnotify v1.6.0
 	github.com/minio/selfupdate v0.6.0
+	github.com/nyaosorg/go-windows-su v0.2.1
 	github.com/ubuntu/gowsl v0.0.0-20230615094051-94945650cc1e
-	github.com/wailsapp/wails/v2 v2.5.1
+	github.com/wailsapp/wails/v2 v2.6.0
 )
 
 require (
 	aead.dev/minisign v0.2.0 // indirect
 	github.com/bep/debounce v1.2.1 // indirect
-	github.com/fsnotify/fsnotify v1.6.0
 	github.com/go-ole/go-ole v1.2.6 // indirect
 	github.com/google/uuid v1.3.0 // indirect
 	github.com/jchv/go-winloader v0.0.0-20210711035445-715c2860da7e // indirect
@@ -22,8 +23,7 @@ require (
 	github.com/leaanthony/gosod v1.0.3 // indirect
 	github.com/leaanthony/slicer v1.6.0 // indirect
 	github.com/mattn/go-colorable v0.1.13 // indirect
-	github.com/mattn/go-isatty v0.0.18 // indirect
-	github.com/nyaosorg/go-windows-su v0.2.1
+	github.com/mattn/go-isatty v0.0.19 // indirect
 	github.com/pkg/browser v0.0.0-20210911075715-681adbf594b8 // indirect
 	github.com/pkg/errors v0.9.1 // indirect
 	github.com/rivo/uniseg v0.4.4 // indirect
@@ -33,9 +33,10 @@ require (
 	github.com/ubuntu/decorate v0.0.0-20230125165522-2d5b0a9bb117 // indirect
 	github.com/valyala/bytebufferpool v1.0.0 // indirect
 	github.com/valyala/fasttemplate v1.2.2 // indirect
+	github.com/wailsapp/go-webview2 v1.0.1 // indirect
 	github.com/wailsapp/mimetype v1.4.1 // indirect
 	golang.org/x/crypto v0.9.0 // indirect
-	golang.org/x/exp v0.0.0-20230515195305-f3d0a9c9a5cc // indirect
+	golang.org/x/exp v0.0.0-20230522175609-2e198f4a06a1 // indirect
 	golang.org/x/net v0.10.0 // indirect
 	golang.org/x/sys v0.9.0 // indirect
 	golang.org/x/text v0.9.0 // indirect

go.sum

@@ -36,8 +36,8 @@ github.com/mattn/go-colorable v0.1.13 h1:fFA4WZxdEF4tXPZVKMLwD8oUnCTTo08duU7wxec
 github.com/mattn/go-colorable v0.1.13/go.mod h1:7S9/ev0klgBDR4GtXTXX8a3vIGJpMovkB8vQcUbaXHg=
 github.com/mattn/go-isatty v0.0.14/go.mod h1:7GGIvUiUoEMVVmxf/4nioHXj79iQHKdU27kJ6hsGG94=
 github.com/mattn/go-isatty v0.0.16/go.mod h1:kYGgaQfpe5nmfYZH+SKPsOc2e4SrIfOl2e/yFXSvRLM=
-github.com/mattn/go-isatty v0.0.18 h1:DOKFKCQ7FNG2L1rbrmstDN4QVRdS89Nkh85u68Uwp98=
-github.com/mattn/go-isatty v0.0.18/go.mod h1:W+V8PltTTMOvKvAeJH7IuucS94S2C6jfK/D7dTCTo3Y=
+github.com/mattn/go-isatty v0.0.19 h1:JITubQf0MOLdlGRuRq+jtsDlekdYPia9ZFsB8h/APPA=
+github.com/mattn/go-isatty v0.0.19/go.mod h1:W+V8PltTTMOvKvAeJH7IuucS94S2C6jfK/D7dTCTo3Y=
 github.com/minio/selfupdate v0.6.0 h1:i76PgT0K5xO9+hjzKcacQtO7+MjJ4JKA8Ak8XQ9DDwU=
 github.com/minio/selfupdate v0.6.0/go.mod h1:bO02GTIPCMQFTEvE5h4DjYB58bCoZ35XLeBf0buTDdM=
 github.com/nyaosorg/go-windows-su v0.2.1 h1:5V0XavLyjOqPUp7psxxCvBISaneU4XmFPSMlejSl5sc=
@@ -69,17 +69,19 @@ github.com/valyala/bytebufferpool v1.0.0/go.mod h1:6bBcMArwyJ5K/AmCkWv1jt77kVWyC
 github.com/valyala/fasttemplate v1.2.1/go.mod h1:KHLXt3tVN2HBp8eijSv/kGJopbvo7S+qRAEEKiv+SiQ=
 github.com/valyala/fasttemplate v1.2.2 h1:lxLXG0uE3Qnshl9QyaK6XJxMXlQZELvChBOCmQD0Loo=
 github.com/valyala/fasttemplate v1.2.2/go.mod h1:KHLXt3tVN2HBp8eijSv/kGJopbvo7S+qRAEEKiv+SiQ=
+github.com/wailsapp/go-webview2 v1.0.1 h1:dEJIeEApW/MhO2tTMISZBFZPuW7kwrFA1NtgFB1z1II=
+github.com/wailsapp/go-webview2 v1.0.1/go.mod h1:Uk2BePfCRzttBBjFrBmqKGJd41P6QIHeV9kTgIeOZNo=
 github.com/wailsapp/mimetype v1.4.1 h1:pQN9ycO7uo4vsUUuPeHEYoUkLVkaRntMnHJxVwYhwHs=
 github.com/wailsapp/mimetype v1.4.1/go.mod h1:9aV5k31bBOv5z6u+QP8TltzvNGJPmNJD4XlAL3U+j3o=
-github.com/wailsapp/wails/v2 v2.5.1 h1:mfG+2kWqQXYOwdgI43HEILjOZDXbk5woPYI3jP2b+js=
-github.com/wailsapp/wails/v2 v2.5.1/go.mod h1:jbOZbcr/zm79PxXxAjP8UoVlDd9wLW3uDs+isIthDfs=
+github.com/wailsapp/wails/v2 v2.6.0 h1:EyH0zR/EO6dDiqNy8qU5spaXDfkluiq77xrkabPYD4c=
+github.com/wailsapp/wails/v2 v2.6.0/go.mod h1:WBG9KKWuw0FKfoepBrr/vRlyTmHaMibWesK3yz6nNiM=
 golang.org/x/crypto v0.0.0-20190308221718-c2843e01d9a2/go.mod h1:djNgcEr1/C05ACkg1iLfiJU5Ep61QUkGW8qpdssI0+w=
 golang.org/x/crypto v0.0.0-20210220033148-5ea612d1eb83/go.mod h1:jdWPYTVW3xRLrWPugEBEK3UY2ZEsg3UU495nc5E+M+I=
 golang.org/x/crypto v0.0.0-20211209193657-4570a0811e8b/go.mod h1:IxCIyHEi3zRg3s0A5j5BB6A9Jmi73HwBIUl50j+osU4=
 golang.org/x/crypto v0.9.0 h1:LF6fAI+IutBocDJ2OT0Q1g8plpYljMZ4+lty+dsqw3g=
 golang.org/x/crypto v0.9.0/go.mod h1:yrmDGqONDYtNj3tH8X9dzUun2m2lzPa9ngI6/RUPGR0=
-golang.org/x/exp v0.0.0-20230515195305-f3d0a9c9a5cc h1:mCRnTeVUjcrhlRmO0VK8a6k6Rrf6TF9htwo2pJVSjIU=
-golang.org/x/exp v0.0.0-20230515195305-f3d0a9c9a5cc/go.mod h1:V1LtkGg67GoY2N1AnLN78QLrzxkLyJw7RJb1gzOOz9w=
+golang.org/x/exp v0.0.0-20230522175609-2e198f4a06a1 h1:k/i9J1pBpvlfR+9QsetwPyERsqu1GIbi967PQMq3Ivc=
+golang.org/x/exp v0.0.0-20230522175609-2e198f4a06a1/go.mod h1:V1LtkGg67GoY2N1AnLN78QLrzxkLyJw7RJb1gzOOz9w=
 golang.org/x/net v0.0.0-20190404232315-eb5bcb51f2a3/go.mod h1:t9HGtf8HONx5eT2rtn7q6eTqICYqUVnKs3thJo3Qplg=
 golang.org/x/net v0.0.0-20210505024714-0287a6fb4125/go.mod h1:9nx3DQGgdP8bBQD5qxJ1jj9UTztislL4KSBs9R2vV5Y=
 golang.org/x/net v0.0.0-20211112202133-69e39bad7dc2/go.mod h1:9nx3DQGgdP8bBQD5qxJ1jj9UTztislL4KSBs9R2vV5Y=

main.go

@@ -98,6 +98,7 @@ func main() {
 		Height:    680,
 		MinWidth:  375,
 		MinHeight: 640,
+		EnableDefaultContextMenu: true,
 		Windows: &windows.Options{
 			ZoomFactor:           zoomFactor,
 			IsZoomControlEnabled: true,

manifest.json

@@ -1,5 +1,5 @@
 {
-  "version": "1.4.4",
+  "version": "1.4.6",
   "introduction": {
     "en": "RWKV is an open-source, commercially usable large language model with high flexibility and great potential for development.\n### About This Tool\nThis tool aims to lower the barrier of entry for using large language models, making it accessible to everyone. It provides fully automated dependency and model management. You simply need to click and run, following the instructions, to deploy a local large language model. The tool itself is very compact and only requires a single executable file for one-click deployment.\nAdditionally, this tool offers an interface that is fully compatible with the OpenAI API. This means you can use any ChatGPT client as a client for RWKV, enabling capability expansion beyond just chat functionality.\n### Preset Configuration Rules at the Bottom\nThis tool comes with a series of preset configurations to reduce complexity. The naming rules for each configuration represent the following in order: device - required VRAM/memory - model size - model language.\nFor example, \"GPU-8G-3B-EN\" indicates that this configuration is for a graphics card with 8GB of VRAM, a model size of 3 billion parameters, and it uses an English language model.\nLarger model sizes have higher performance and VRAM requirements. Among configurations with the same model size, those with higher VRAM usage will have faster runtime.\nFor example, if you have 12GB of VRAM but running the \"GPU-12G-7B-EN\" configuration is slow, you can downgrade to \"GPU-8G-3B-EN\" for a significant speed improvement.\n### About RWKV\nRWKV is an RNN with Transformer-level LLM performance, which can also be directly trained like a GPT transformer (parallelizable). And it's 100% attention-free. You only need the hidden state at position t to compute the state at position t+1. You can use the \"GPT\" mode to quickly compute the hidden state for the \"RNN\" mode.<br/>So it's combining the best of RNN and transformer - great performance, fast inference, saves VRAM, fast training, \"infinite\" ctx_len, and free sentence embedding (using the final hidden state).",
     "zh": "RWKV是一个开源且允许商用的大语言模型，灵活性很高且极具发展潜力。\n### 关于本工具\n本工具旨在降低大语言模型的使用门槛，做到人人可用，本工具提供了全自动化的依赖和模型管理，你只需要直接点击运行，跟随引导，即可完成本地大语言模型的部署，工具本身体积极小，只需要一个exe即可完成一键部署。\n此外，本工具提供了与OpenAI API完全兼容的接口，这意味着你可以把任意ChatGPT客户端用作RWKV的客户端，实现能力拓展，而不局限于聊天。\n### 底部的预设配置规则\n本工具内置了一系列预设配置，以降低使用难度，每个配置名的规则，依次代表着：设备-所需显存/内存-模型规模-模型语言。\n例如，GPU-8G-3B-CN，表示该配置用于显卡，需要8G显存，模型规模为30亿参数，使用的是中文模型。\n模型规模越大，性能要求越高，显存要求也越高，而同样模型规模的配置中，显存占用越高的，运行速度越快。\n例如当你有12G显存，但运行GPU-12G-7B-CN配置速度比较慢，可降级成GPU-8G-3B-CN，将会大幅提速。\n### 关于RWKV\nRWKV是具有Transformer级别LLM性能的RNN，也可以像GPT Transformer一样直接进行训练（可并行化）。而且它是100% attention-free的。你只需在位置t处获得隐藏状态即可计算位置t + 1处的状态。你可以使用“GPT”模式快速计算用于“RNN”模式的隐藏状态。\n因此，它将RNN和Transformer的优点结合起来 - 高性能、快速推理、节省显存、快速训练、“无限”上下文长度以及免费的语句嵌入（使用最终隐藏状态）。"