Merge a8dc94ebb05b5f26c757a29bd32ea6616b6a9aea into 11aa78bd9bda8b53047cfcae03abf7ca94d27391

修复环境变量可能不为str的问题

.github/build_windows_packages.ps1

@@ -24,6 +24,29 @@ if (-not [string]::IsNullOrWhiteSpace($suffix)) {
 
 $pkgName = "$pkgName-$cuda"
 
+$SevenZipPath = "C:\Program Files\7-Zip\7z.exe"
+$SevenZipDir  = Split-Path $SevenZipPath
+$CodecsDir    = Join-Path $SevenZipDir "Codecs"
+
+$Url = "https://github.com/mcmilk/7-Zip-zstd/releases/download/v25.01-v1.5.7-R1/Codecs-x64.7z"
+
+$TempArchive = "$env:TEMP\Codecs-x64.7z"
+
+Write-Host "Downloading 7-Zip Zstd plugin..."
+Invoke-WebRequest -Uri $Url -OutFile $TempArchive
+
+if (-not (Test-Path $CodecsDir)) {
+    New-Item -Path $CodecsDir -ItemType Directory | Out-Null
+    Write-Host "Created Codecs directory: $CodecsDir"
+}
+
+Write-Host "Extracting plugin..."
+& $SevenZipPath x $TempArchive "-o$CodecsDir" -y | Out-Null
+
+Remove-Item $TempArchive -Force
+
+Write-Host "Patch complete. Installed plugins in $CodecsDir"
+
 $baseHF = "https://huggingface.co/XXXXRT/GPT-SoVITS-Pretrained/resolve/main"
 $PRETRAINED_URL = "$baseHF/pretrained_models.zip"
 $G2PW_URL = "$baseHF/G2PWModel.zip"
@@ -118,14 +141,14 @@ Write-Host "[INFO] Installing PyTorch..."
 
 switch ($cuda) {
     "cu126" {
-        & ".\runtime\python.exe" -m pip install psutil ninja packaging wheel "setuptools>=42" --no-warn-script-location
-        & ".\runtime\python.exe" -m pip install torch --index-url https://download.pytorch.org/whl/cu126 --no-warn-script-location
-        & ".\runtime\python.exe" -m pip install flash-attn -i https://xxxxrt666.github.io/PIP-Index/ --no-build-isolation
+        & ".\runtime\python.exe" -m pip install psutil ninja packaging wheel "setuptools>=42" --no-warn-script-location --no-cache-dir
+        & ".\runtime\python.exe" -m pip install torch --index-url https://download.pytorch.org/whl/cu126 --no-warn-script-location --no-cache-dir
+        & ".\runtime\python.exe" -m pip install flash-attn -i https://xxxxrt666.github.io/PIP-Index/ --no-build-isolation --no-cache-dir
     }
     "cu128" {
-        & ".\runtime\python.exe" -m pip install psutil ninja packaging wheel "setuptools>=42" --no-warn-script-location
-        & ".\runtime\python.exe" -m pip install torch --index-url https://download.pytorch.org/whl/cu128 --no-warn-script-location
-        & ".\runtime\python.exe" -m pip install flash-attn -i https://xxxxrt666.github.io/PIP-Index/ --no-build-isolation
+        & ".\runtime\python.exe" -m pip install psutil ninja packaging wheel "setuptools>=42" --no-warn-script-location --no-cache-dir
+        & ".\runtime\python.exe" -m pip install torch --index-url https://download.pytorch.org/whl/cu128 --no-warn-script-location --no-cache-dir
+        & ".\runtime\python.exe" -m pip install flash-attn -i https://xxxxrt666.github.io/PIP-Index/ --no-build-isolation --no-cache-dir
     }
     default {
         Write-Error "Unsupported CUDA version: $cuda"
@@ -168,7 +191,7 @@ Copy-Item -Path $curr -Destination $pkgName -Recurse
 $7zPath = "$pkgName.7z"
 $start = Get-Date
 Write-Host "Compress Starting at $start"
-& "C:\Program Files\7-Zip\7z.exe" a -t7z "$7zPath" "$pkgName" -m0=lzma2 -mx=9 -mmt=on -bsp1
+& "C:\Program Files\7-Zip\7z.exe" a -t7z "$7zPath" "$pkgName" -m0=bcj -m1=zstd -mx=22 -mmt=on -bsp1
 $end = Get-Date
 Write-Host "Elapsed time: $($end - $start)"
 Get-ChildItem .

.github/workflows/build_windows_packages.yaml

@@ -28,9 +28,6 @@ jobs:
       PKG_SUFFIX: ${{ github.event.inputs.suffix }}
 
     steps:
-      - name: Checkout
-        uses: actions/checkout@v4
-
       - name: Install Windows CUDA 12.9
         uses: Jimver/cuda-toolkit@v0.2.24
         id: cuda-toolkit-win-129
@@ -39,6 +36,17 @@ jobs:
           method: "network"
           sub-packages: '["nvcc", "cudart", "visual_studio_integration"]'
 
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Cache pip
+        uses: actions/cache@v4
+        with:
+          path: ~\AppData\Local\pip\Cache
+          key: ${{ runner.os }}-pip-${{ hashFiles('**/requirements.txt') }}
+          restore-keys: |
+            ${{ runner.os }}-pip-
+
       - name: Run Build and Upload Script
         shell: pwsh
         run: |

GPT_SoVITS/Accelerate/MLX/backends/mlx_quantized.py

@@ -1,7 +1,5 @@
 from __future__ import annotations
 
-from typing import cast
-
 import mlx.core as mx
 import mlx.nn as nn
 
@@ -43,7 +41,7 @@ class Attention(AttentionABC):
         return out
 
     def __call__(self, x: Array, input_pos: Array, kv_cache: KVCache | KVCacheQ, cache_idx: Array, attn_mask: Array):
-        bsz, seqlen, _ = cast(tuple[int, ...], x.shape)
+        bsz, seqlen, _ = x.shape
 
         q, k, v = self.in_proj(x).split(3, axis=-1)
 
@@ -69,7 +67,7 @@ class Attention(AttentionABC):
         return attn
 
     # def __call__(self, x: Array, input_pos: Array, kv_cache: KVCache | KVCacheQ, cache_idx: Array, attn_mask: Array):
-    #     bsz, seqlen, _ = cast(tuple[int, ...], x.shape)
+    #     bsz, seqlen, _ = x.shape
 
     #     q, k, v = self.in_proj(x).split(3, axis=-1)
 

GPT_SoVITS/Accelerate/MLX/backends/mlx_static.py

@@ -1,7 +1,5 @@
 from __future__ import annotations
 
-from typing import cast
-
 import mlx.core as mx
 
 from ..structs_mlx import KVCache, KVCacheQ
@@ -22,7 +20,7 @@ class Attention(AttentionABC):
         self.kc_class = KVCacheHND
 
     def __call__(self, x: Array, input_pos: Array, kv_cache: KVCache | KVCacheQ, cache_idx: Array, attn_mask: Array):
-        bsz, seqlen, _ = cast(tuple[int, ...], x.shape)
+        bsz, seqlen, _ = x.shape
 
         q, k, v = self.in_proj(x).split(3, axis=-1)
 

GPT_SoVITS/Accelerate/MLX/backends/mlx_varlen.py

@@ -1,7 +1,5 @@
 from __future__ import annotations
 
-from typing import cast
-
 import mlx.core as mx
 
 from ..structs_mlx import KVCache, KVCacheQ
@@ -22,7 +20,7 @@ class Attention(AttentionABC):
         self.kc_class = KVCacheHND
 
     def __call__(self, x: Array, input_pos: Array, kv_cache: KVCache | KVCacheQ, cache_idx: Array, attn_mask: Array):
-        bsz, seqlen, _ = cast(tuple[int, ...], x.shape)
+        bsz, seqlen, _ = x.shape
 
         q, k, v = self.in_proj(x).split(3, axis=-1)
 

GPT_SoVITS/Accelerate/MLX/sample_funcs_mlx.py

@@ -1,4 +1,4 @@
-from typing import Protocol, cast
+from typing import Protocol
 
 import mlx.core as mx
 
@@ -33,7 +33,7 @@ class sample_naive(SampleProtocolMLX):
             return mx.argmax(probs, axis=-1, keepdims=True).astype(mx.int32)
 
         if repetition_penalty != 1.0:
-            batch_idx = mx.arange(cast(tuple[int, ...], previous_tokens.shape)[0])
+            batch_idx = mx.arange(previous_tokens.shape[0])
             previous_tokens = previous_tokens.astype(mx.int64)
             selected_logists = logits[batch_idx, previous_tokens]
             selected_logists = mx.where(
@@ -48,7 +48,7 @@ class sample_naive(SampleProtocolMLX):
             sorted_indices_to_remove = cum_probs > top_p
             sorted_indices_to_remove[:, -1] = False
             indices_to_remove = mx.zeros_like(logits).astype(mx.bool_)
-            batch_indices = mx.arange(cast(tuple[int, ...], logits.shape)[0])[:, None]
+            batch_indices = mx.arange(logits.shape[0])[:, None]
             indices_to_remove[batch_indices, sorted_indices] = sorted_indices_to_remove
             logits = mx.where(indices_to_remove, -mx.inf, logits)
 
@@ -59,7 +59,7 @@ class sample_naive(SampleProtocolMLX):
         pivot = mx.expand_dims(v[:, 0], -1)
         logits = mx.where(logits < pivot, -mx.inf, logits)
 
-        gumbel_noise = mx.random.gumbel(shape=cast(tuple[int, ...], logits.shape), dtype=logits.dtype)
+        gumbel_noise = mx.random.gumbel(shape=logits.shape, dtype=logits.dtype)
         idx_next = mx.argmax(logits + gumbel_noise, axis=-1, keepdims=True).astype(mx.int32)
 
         return idx_next

GPT_SoVITS/Accelerate/MLX/structs_mlx.py

@@ -5,7 +5,7 @@ Modified From https://github.com/XXXXRT666/GPT-SoVITS
 from __future__ import annotations
 
 from dataclasses import dataclass
-from typing import List, MutableSequence, Protocol, TypeAlias, cast
+from typing import List, MutableSequence, Protocol, TypeAlias
 
 import mlx.core as mx
 import torch
@@ -32,10 +32,10 @@ class T2SRequestMLX:
 
     @classmethod
     def from_torch(cls, request: T2SRequest) -> T2SRequestMLX:
-        x = list(map(lambda tensor: mx.array(tensor.cpu()), request.x))
-        x_lens = mx.array(request.x_lens.cpu())
-        prompts = mx.array(request.prompts.cpu())
-        bert_feature = list(map(lambda tensor: mx.array(tensor.cpu()), request.bert_feature))
+        x = list(map(lambda tensor: mx.array(tensor.cpu()), request.x))  # type: ignore
+        x_lens = mx.array(request.x_lens.cpu())  # type: ignore
+        prompts = mx.array(request.prompts.cpu())  # type: ignore
+        bert_feature = list(map(lambda tensor: mx.array(tensor.cpu()), request.bert_feature))  # type: ignore
 
         return cls(
             x,
@@ -99,7 +99,7 @@ class T2SSessionMLX:
             self.dtype = dtype
 
             bsz = len(request.x)
-            y_len: int = cast(tuple[int, ...], request.prompts.shape)[-1]
+            y_len: int = request.prompts.shape[-1]
             self.bsz = bsz
             self.y_len = y_len
 
@@ -111,10 +111,10 @@ class T2SSessionMLX:
             self.x = [i.astype(mx.int32) for i in request.x]
             self.x_lens = request.x_lens.astype(mx.int32)
             self.y = mx.zeros((bsz, decoder.max_seq_length)).astype(mx.int32)
-            self.y[:, : cast(tuple[int, ...], request.prompts.shape)[-1]] = request.prompts.astype(mx.int32)
+            self.y[:, : request.prompts.shape[-1]] = request.prompts.astype(mx.int32)
             self.bert_feature = [i.astype(dtype) for i in request.bert_feature]
 
-            self.prefill_len = self.x_lens + cast(tuple[int, ...], request.prompts.shape)[1]
+            self.prefill_len = self.x_lens + request.prompts.shape[1]
 
             self.input_pos = mx.zeros_like(self.prefill_len)
             self.input_pos += self.prefill_len

GPT_SoVITS/Accelerate/MLX/t2s_engine_mlx.py

@@ -140,7 +140,7 @@ class T2SEngine(T2SEngineProtocol):
                         logger.error("Bad Zero Prediction")
                     else:
                         logger.info(
-                            f"T2S Decoding EOS {session.prefill_len.tolist().__str__().strip('[]')} -> {[cast(tuple[int, ...], i.shape)[-1] for i in session.y_results].__str__().strip('[]')}"
+                            f"T2S Decoding EOS {session.prefill_len.tolist().__str__().strip('[]')} -> {[i.shape[-1] for i in session.y_results].__str__().strip('[]')}"
                         )
                         logger.info(f"Infer Speed: {(idx - 1) / (time.perf_counter() - t1):.2f} token/s")
                         infer_time = time.perf_counter() - t1
@@ -199,7 +199,7 @@ class T2SEngine(T2SEngineProtocol):
                 .replace("norm1", "attention_norm")
                 .replace("norm2", "ffn_norm")
             )
-            value_mlx = mx.array(value)
+            value_mlx = mx.array(value)  # type: ignore
             state_dict_mlx.append((key, value_mlx))
         return state_dict_mlx
 

GPT_SoVITS/Accelerate/MLX/t2s_model_abc.py

@@ -2,7 +2,7 @@ from __future__ import annotations
 
 import math
 from abc import ABC, abstractmethod
-from typing import MutableSequence, cast
+from typing import MutableSequence
 
 import mlx.core as mx
 import mlx.nn as nn
@@ -85,7 +85,7 @@ class SinePositionalEmbedding(nn.Module):
             embedded_x (Array): [batch_size, 1, embed_dim]
         """
 
-        batch_size = cast(tuple[int, ...], x.shape)[0]
+        batch_size = x.shape[0]
         pe_values = self._pe[mx.arange(batch_size), input_pos - 1]  # (batch_size, embed_dim)
 
         return x * self.x_scale + self.alpha * mx.expand_dims(pe_values, 1)  # (batch_size, 1, embed_dim)
@@ -98,7 +98,7 @@ class SinePositionalEmbedding(nn.Module):
         Returns:
             embedded_x (Array): [batch_size, seq_len, embed_dim]
         """
-        pe_values = self._pe[:, : cast(tuple[int, ...], x.shape)[-2]]
+        pe_values = self._pe[:, : x.shape[-2]]
         return x * self.x_scale + self.alpha * pe_values
 
 
@@ -129,8 +129,8 @@ class KVCacheHND(KVCacheProtocol):
         assert len(kv_cache) == 2
         k_cache, v_cache = kv_cache
 
-        k_cache[..., : cast(tuple[int, ...], k_val.shape)[1], :] = k_val.swapaxes(1, 2)
-        v_cache[..., : cast(tuple[int, ...], v_val.shape)[1], :] = v_val.swapaxes(1, 2)
+        k_cache[..., : k_val.shape[1], :] = k_val.swapaxes(1, 2)
+        v_cache[..., : v_val.shape[1], :] = v_val.swapaxes(1, 2)
 
     @staticmethod
     def init_cache(batch_size: int, max_seq_length: int, n_heads: int, head_dim: int, dtype: mx.Dtype) -> KVCache:
@@ -207,7 +207,7 @@ class KVCacheHNDQuantized(KVCacheProtocol):
         assert len(kv_cache) == 3
         (k_q_out, k_s_out, k_b_out), (v_q_out, v_s_out, v_b_out), (group_size, bits) = kv_cache
 
-        S = cast(tuple[int, ...], k_val.shape)[1]
+        S = k_val.shape[1]
 
         k_sw = k_val.swapaxes(1, 2)
         v_sw = v_val.swapaxes(1, 2)
@@ -276,7 +276,7 @@ class AttentionABC(ABC, nn.Module):
     ) -> Array: ...
 
     def prefill(self, x: Array, kv_cache: KVCache | KVCacheQ, attn_mask: Array):
-        bsz, seqlen, _ = cast(tuple[int, ...], x.shape)
+        bsz, seqlen, _ = x.shape
 
         q, k, v = self.in_proj(x).split(3, axis=-1)
 
@@ -481,15 +481,13 @@ class T2SDecoderABC(nn.Module, T2SDecoderProtocol):
         y: Array,
         bert_features: list[Array],
     ):
-        x_len: list[int] = [cast(tuple[int, ...], i.shape)[0] for i in x]
+        x_len: list[int] = [i.shape[0] for i in x]
         x_len_max = max(x_len)
-        xy_pos = mx.zeros((len(x), x_len_max + cast(tuple[int, ...], y.shape)[1], self.embedding_dim)).astype(
-            bert_features[0].dtype
-        )
+        xy_pos = mx.zeros((len(x), x_len_max + y.shape[1], self.embedding_dim)).astype(bert_features[0].dtype)
 
         bert_features = list(map(lambda x: x.swapaxes(0, 1), bert_features))
 
-        y_len = cast(tuple[int, ...], y.shape)[1]
+        y_len = y.shape[1]
         y_emb = self.ar_audio_embedding(y)
         y_pos = self.ar_audio_position.prefill(y_emb)
 

GPT_SoVITS/inference_webui.py

@@ -7,7 +7,7 @@ import traceback
 import warnings
 from functools import partial
 from pathlib import Path
-from time import time as ttime
+from time import perf_counter as ttime
 from typing import Any
 
 import gradio as gr
@@ -69,7 +69,6 @@ def set_high_priority():
     p = psutil.Process(os.getpid())
     with contextlib.suppress(psutil.AccessDenied):
         p.nice(psutil.HIGH_PRIORITY_CLASS)
-        print("已将进程优先级设为 High")
 
 
 _LANG_RE = re.compile(r"^[a-z]{2}[_-][A-Z]{2}$")
@@ -723,7 +722,6 @@ def get_tts_wav(
             prompt_text += "。" if prompt_language != "en" else "."
         print(">>", i18n("实际输入的参考文本:"), prompt_text)
     text = text.strip("\n")
-    # if (text[0] not in splits and len(get_first(text)) < 4): text = "。" + text if text_language != "en" else "." + text
 
     print(">>", i18n("实际输入的目标文本:"), text)
     zero_wav = np.zeros(
@@ -813,7 +811,7 @@ def get_tts_wav(
                 top_p=top_p,
                 temperature=temperature,
                 early_stop_num=1500,
-                use_cuda_graph=torch.cuda.is_available(),
+                use_cuda_graph=torch.cuda.is_available(),  # Try to use CUDA Graph for all backend, fallback to normal if not applicapble
                 # debug=True,
             )
             assert t2s_engine
@@ -1300,7 +1298,7 @@ with gr.Blocks(title="GPT-SoVITS WebUI", analytics_enabled=False, js=js, css=css
 
 if __name__ == "__main__":
     set_high_priority()
-    app.queue(api_open=False, default_concurrency_limit=512, max_size=1024).launch(
+    app.queue(api_open=False, default_concurrency_limit=1, max_size=1024).launch(
         server_name="0.0.0.0",
         inbrowser=True,
         share=is_share,