feat: support speech2text and tts for gpustack

api/configs/packaging/__init__.py

@@ -9,7 +9,7 @@ class PackagingInfo(BaseSettings):
 
     CURRENT_VERSION: str = Field(
         description="Dify version",
-        default="0.15.0",
+        default="0.14.2",
     )
 
     COMMIT_SHA: str = Field(

api/controllers/service_api/wraps.py

@@ -1,5 +1,5 @@
 from collections.abc import Callable
-from datetime import UTC, datetime, timedelta
+from datetime import UTC, datetime
 from enum import Enum
 from functools import wraps
 from typing import Optional
@@ -8,8 +8,6 @@ from flask import current_app, request
 from flask_login import user_logged_in  # type: ignore
 from flask_restful import Resource  # type: ignore
 from pydantic import BaseModel
-from sqlalchemy import select, update
-from sqlalchemy.orm import Session
 from werkzeug.exceptions import Forbidden, Unauthorized
 
 from extensions.ext_database import db
@@ -176,7 +174,7 @@ def validate_dataset_token(view=None):
     return decorator
 
 
-def validate_and_get_api_token(scope: str | None = None):
+def validate_and_get_api_token(scope=None):
     """
     Validate and get API token.
     """
@@ -190,25 +188,20 @@ def validate_and_get_api_token(scope: str | None = None):
     if auth_scheme != "bearer":
         raise Unauthorized("Authorization scheme must be 'Bearer'")
 
-    current_time = datetime.now(UTC).replace(tzinfo=None)
-    cutoff_time = current_time - timedelta(minutes=1)
-    with Session(db.engine, expire_on_commit=False) as session:
-        update_stmt = (
-            update(ApiToken)
-            .where(ApiToken.token == auth_token, ApiToken.last_used_at < cutoff_time, ApiToken.type == scope)
-            .values(last_used_at=current_time)
-            .returning(ApiToken)
+    api_token = (
+        db.session.query(ApiToken)
+        .filter(
+            ApiToken.token == auth_token,
+            ApiToken.type == scope,
         )
-        result = session.execute(update_stmt)
-        api_token = result.scalar_one_or_none()
+        .first()
+    )
 
-        if not api_token:
-            stmt = select(ApiToken).where(ApiToken.token == auth_token, ApiToken.type == scope)
-            api_token = session.scalar(stmt)
-            if not api_token:
-                raise Unauthorized("Access token is invalid")
-        else:
-            session.commit()
+    if not api_token:
+        raise Unauthorized("Access token is invalid")
+
+    api_token.last_used_at = datetime.now(UTC).replace(tzinfo=None)
+    db.session.commit()
 
     return api_token
 

api/core/model_runtime/model_providers/__base/tokenizers/gpt2_tokenzier.py

@@ -1,10 +1,13 @@
+from concurrent.futures import ProcessPoolExecutor
+from os.path import abspath, dirname, join
 from threading import Lock
-from typing import Any
+from typing import Any, cast
 
-import tiktoken
+from transformers import GPT2Tokenizer as TransformerGPT2Tokenizer  # type: ignore
 
 _tokenizer: Any = None
 _lock = Lock()
+_executor = ProcessPoolExecutor(max_workers=1)
 
 
 class GPT2Tokenizer:
@@ -14,28 +17,22 @@ class GPT2Tokenizer:
         use gpt2 tokenizer to get num tokens
         """
         _tokenizer = GPT2Tokenizer.get_encoder()
-        tokens = _tokenizer.encode(text)
+        tokens = _tokenizer.encode(text, verbose=False)
         return len(tokens)
 
     @staticmethod
     def get_num_tokens(text: str) -> int:
-        # Because this process needs more cpu resource, we turn this back before we find a better way to handle it.
-        #
-        # future = _executor.submit(GPT2Tokenizer._get_num_tokens_by_gpt2, text)
-        # result = future.result()
-        # return cast(int, result)
-        return GPT2Tokenizer._get_num_tokens_by_gpt2(text)
+        future = _executor.submit(GPT2Tokenizer._get_num_tokens_by_gpt2, text)
+        result = future.result()
+        return cast(int, result)
 
     @staticmethod
     def get_encoder() -> Any:
         global _tokenizer, _lock
         with _lock:
             if _tokenizer is None:
-                # Try to use tiktoken to get the tokenizer because it is faster
-                #
-                _tokenizer = tiktoken.get_encoding("gpt2")
-                # base_path = abspath(__file__)
-                # gpt2_tokenizer_path = join(dirname(base_path), "gpt2")
-                # _tokenizer = TransformerGPT2Tokenizer.from_pretrained(gpt2_tokenizer_path)
+                base_path = abspath(__file__)
+                gpt2_tokenizer_path = join(dirname(base_path), "gpt2")
+                _tokenizer = TransformerGPT2Tokenizer.from_pretrained(gpt2_tokenizer_path)
 
             return _tokenizer

api/core/model_runtime/model_providers/openai_api_compatible/llm/llm.py

@@ -377,10 +377,7 @@ class OAIAPICompatLargeLanguageModel(_CommonOaiApiCompat, LargeLanguageModel):
                 for tool in tools:
                     formatted_tools.append(helper.dump_model(PromptMessageFunction(function=tool)))
 
-                if prompt_messages[-1].role.value == "tool":
-                    data["tools"] = None
-                else:
-                    data["tools"] = formatted_tools
+                data["tools"] = formatted_tools
 
         if stop:
             data["stop"] = stop

api/core/model_runtime/model_providers/siliconflow/llm/_position.yaml

@@ -7,8 +7,6 @@
 - Qwen/Qwen2.5-Coder-7B-Instruct
 - Qwen/Qwen2-VL-72B-Instruct
 - Qwen/Qwen2-1.5B-Instruct
-- Qwen/Qwen2.5-72B-Instruct-128K
-- Vendor-A/Qwen/Qwen2.5-72B-Instruct
 - Pro/Qwen/Qwen2-VL-7B-Instruct
 - OpenGVLab/InternVL2-26B
 - Pro/OpenGVLab/InternVL2-8B

api/core/model_runtime/model_providers/siliconflow/llm/qwen2.5-72b-instruct-128k.yaml

@@ -1,51 +0,0 @@
-model: Qwen/Qwen2.5-72B-Instruct-128K
-label:
-  en_US: Qwen/Qwen2.5-72B-Instruct-128K
-model_type: llm
-features:
-  - agent-thought
-model_properties:
-  mode: chat
-  context_size: 131072
-parameter_rules:
-  - name: temperature
-    use_template: temperature
-  - name: max_tokens
-    use_template: max_tokens
-    type: int
-    default: 512
-    min: 1
-    max: 4096
-    help:
-      zh_Hans: 指定生成结果长度的上限。如果生成结果截断，可以调大该参数。
-      en_US: Specifies the upper limit on the length of generated results. If the generated results are truncated, you can increase this parameter.
-  - name: top_p
-    use_template: top_p
-  - name: top_k
-    label:
-      zh_Hans: 取样数量
-      en_US: Top k
-    type: int
-    help:
-      zh_Hans: 仅从每个后续标记的前 K 个选项中采样。
-      en_US: Only sample from the top K options for each subsequent token.
-    required: false
-  - name: frequency_penalty
-    use_template: frequency_penalty
-  - name: response_format
-    label:
-      zh_Hans: 回复格式
-      en_US: Response Format
-    type: string
-    help:
-      zh_Hans: 指定模型必须输出的格式
-      en_US: specifying the format that the model must output
-    required: false
-    options:
-      - text
-      - json_object
-pricing:
-  input: '4.13'
-  output: '4.13'
-  unit: '0.000001'
-  currency: RMB

api/core/model_runtime/model_providers/siliconflow/llm/qwen2.5-72b-instruct-vendorA.yaml

@@ -1,51 +0,0 @@
-model: Vendor-A/Qwen/Qwen2.5-72B-Instruct
-label:
-  en_US: Vendor-A/Qwen/Qwen2.5-72B-Instruct
-model_type: llm
-features:
-  - agent-thought
-model_properties:
-  mode: chat
-  context_size: 32768
-parameter_rules:
-  - name: temperature
-    use_template: temperature
-  - name: max_tokens
-    use_template: max_tokens
-    type: int
-    default: 512
-    min: 1
-    max: 4096
-    help:
-      zh_Hans: 指定生成结果长度的上限。如果生成结果截断，可以调大该参数。
-      en_US: Specifies the upper limit on the length of generated results. If the generated results are truncated, you can increase this parameter.
-  - name: top_p
-    use_template: top_p
-  - name: top_k
-    label:
-      zh_Hans: 取样数量
-      en_US: Top k
-    type: int
-    help:
-      zh_Hans: 仅从每个后续标记的前 K 个选项中采样。
-      en_US: Only sample from the top K options for each subsequent token.
-    required: false
-  - name: frequency_penalty
-    use_template: frequency_penalty
-  - name: response_format
-    label:
-      zh_Hans: 回复格式
-      en_US: Response Format
-    type: string
-    help:
-      zh_Hans: 指定模型必须输出的格式
-      en_US: specifying the format that the model must output
-    required: false
-    options:
-      - text
-      - json_object
-pricing:
-  input: '1.00'
-  output: '1.00'
-  unit: '0.000001'
-  currency: RMB

api/core/model_runtime/model_providers/siliconflow/llm/qwen2.5-72b-instruct.yaml

@@ -15,7 +15,7 @@ parameter_rules:
     type: int
     default: 512
     min: 1
-    max: 4096
+    max: 8192
     help:
       zh_Hans: 指定生成结果长度的上限。如果生成结果截断，可以调大该参数。
       en_US: Specifies the upper limit on the length of generated results. If the generated results are truncated, you can increase this parameter.

api/core/model_runtime/model_providers/siliconflow/tts/fish-speech-1.5.yaml

@@ -1,37 +0,0 @@
-model: fishaudio/fish-speech-1.5
-model_type: tts
-model_properties:
-  default_voice: 'fishaudio/fish-speech-1.5:alex'
-  voices:
-    - mode: "fishaudio/fish-speech-1.5:alex"
-      name: "Alex（男声）"
-      language: [ "zh-Hans", "en-US" ]
-    - mode: "fishaudio/fish-speech-1.5:benjamin"
-      name: "Benjamin（男声）"
-      language: [ "zh-Hans", "en-US" ]
-    - mode: "fishaudio/fish-speech-1.5:charles"
-      name: "Charles（男声）"
-      language: [ "zh-Hans", "en-US" ]
-    - mode: "fishaudio/fish-speech-1.5:david"
-      name: "David（男声）"
-      language: [ "zh-Hans", "en-US" ]
-    - mode: "fishaudio/fish-speech-1.5:anna"
-      name: "Anna（女声）"
-      language: [ "zh-Hans", "en-US" ]
-    - mode: "fishaudio/fish-speech-1.5:bella"
-      name: "Bella（女声）"
-      language: [ "zh-Hans", "en-US" ]
-    - mode: "fishaudio/fish-speech-1.5:claire"
-      name: "Claire（女声）"
-      language: [ "zh-Hans", "en-US" ]
-    - mode: "fishaudio/fish-speech-1.5:diana"
-      name: "Diana（女声）"
-      language: [ "zh-Hans", "en-US" ]
-  audio_type: 'mp3'
-  max_workers: 5
-  # stream: false
-pricing:
-  input: '0.015'
-  output: '0'
-  unit: '0.001'
-  currency: RMB

api/core/workflow/nodes/document_extractor/node.py

@@ -2,7 +2,6 @@ import csv
 import io
 import json
 import logging
-import operator
 import os
 import tempfile
 from typing import cast
@@ -11,8 +10,6 @@ import docx
 import pandas as pd
 import pypdfium2  # type: ignore
 import yaml  # type: ignore
-from docx.table import Table
-from docx.text.paragraph import Paragraph
 
 from configs import dify_config
 from core.file import File, FileTransferMethod, file_manager
@@ -192,56 +189,35 @@ def _extract_text_from_doc(file_content: bytes) -> str:
         doc_file = io.BytesIO(file_content)
         doc = docx.Document(doc_file)
         text = []
-
-        # Keep track of paragraph and table positions
-        content_items: list[tuple[int, str, Table | Paragraph]] = []
-
-        # Process paragraphs and tables
-        for i, paragraph in enumerate(doc.paragraphs):
+        # Process paragraphs
+        for paragraph in doc.paragraphs:
             if paragraph.text.strip():
-                content_items.append((i, "paragraph", paragraph))
+                text.append(paragraph.text)
 
-        for i, table in enumerate(doc.tables):
-            content_items.append((i, "table", table))
-
-        # Sort content items based on their original position
-        content_items.sort(key=operator.itemgetter(0))
-
-        # Process sorted content
-        for _, item_type, item in content_items:
-            if item_type == "paragraph":
-                if isinstance(item, Table):
-                    continue
-                text.append(item.text)
-            elif item_type == "table":
-                # Process tables
-                if not isinstance(item, Table):
-                    continue
-                try:
+        # Process tables
+        for table in doc.tables:
+            # Table header
+            try:
+                # table maybe cause errors so ignore it.
+                if len(table.rows) > 0 and table.rows[0].cells is not None:
                     # Check if any cell in the table has text
                     has_content = False
-                    for row in item.rows:
+                    for row in table.rows:
                         if any(cell.text.strip() for cell in row.cells):
                             has_content = True
                             break
 
                     if has_content:
-                        cell_texts = [cell.text.replace("\n", "<br>") for cell in item.rows[0].cells]
-                        markdown_table = f"| {' | '.join(cell_texts)} |\n"
-                        markdown_table += f"| {' | '.join(['---'] * len(item.rows[0].cells))} |\n"
-
-                        for row in item.rows[1:]:
-                            # Replace newlines with <br> in each cell
-                            row_cells = [cell.text.replace("\n", "<br>") for cell in row.cells]
-                            markdown_table += "| " + " | ".join(row_cells) + " |\n"
-
+                        markdown_table = "| " + " | ".join(cell.text for cell in table.rows[0].cells) + " |\n"
+                        markdown_table += "| " + " | ".join(["---"] * len(table.rows[0].cells)) + " |\n"
+                        for row in table.rows[1:]:
+                            markdown_table += "| " + " | ".join(cell.text for cell in row.cells) + " |\n"
                         text.append(markdown_table)
-                except Exception as e:
-                    logger.warning(f"Failed to extract table from DOC/DOCX: {e}")
-                    continue
+            except Exception as e:
+                logger.warning(f"Failed to extract table from DOC/DOCX: {e}")
+                continue
 
         return "\n".join(text)
-
     except Exception as e:
         raise TextExtractionError(f"Failed to extract text from DOC/DOCX: {str(e)}") from e
 

api/docker/entrypoint.sh

@@ -33,7 +33,6 @@ else
       --bind "${DIFY_BIND_ADDRESS:-0.0.0.0}:${DIFY_PORT:-5001}" \
       --workers ${SERVER_WORKER_AMOUNT:-1} \
       --worker-class ${SERVER_WORKER_CLASS:-gevent} \
-      --worker-connections ${SERVER_WORKER_CONNECTIONS:-10} \
       --timeout ${GUNICORN_TIMEOUT:-200} \
       app:app
   fi

api/services/billing_service.py

@@ -1,5 +1,5 @@
 import os
-from typing import Literal, Optional
+from typing import Optional
 
 import httpx
 from tenacity import retry, retry_if_exception_type, stop_before_delay, wait_fixed
@@ -17,6 +17,7 @@ class BillingService:
         params = {"tenant_id": tenant_id}
 
         billing_info = cls._send_request("GET", "/subscription/info", params=params)
+
         return billing_info
 
     @classmethod
@@ -46,13 +47,12 @@ class BillingService:
         retry=retry_if_exception_type(httpx.RequestError),
         reraise=True,
     )
-    def _send_request(cls, method: Literal["GET", "POST", "DELETE"], endpoint: str, json=None, params=None):
+    def _send_request(cls, method, endpoint, json=None, params=None):
         headers = {"Content-Type": "application/json", "Billing-Api-Secret-Key": cls.secret_key}
 
         url = f"{cls.base_url}{endpoint}"
         response = httpx.request(method, url, json=json, params=params, headers=headers)
-        if method == "GET" and response.status_code != httpx.codes.OK:
-            raise ValueError("Unable to retrieve billing information. Please try again later or contact support.")
+
         return response.json()
 
     @staticmethod

docker-legacy/docker-compose.yaml

@@ -2,7 +2,7 @@ version: '3'
 services:
   # API service
   api:
-    image: langgenius/dify-api:0.15.0
+    image: langgenius/dify-api:0.14.2
     restart: always
     environment:
       # Startup mode, 'api' starts the API server.
@@ -227,7 +227,7 @@ services:
   # worker service
   # The Celery worker for processing the queue.
   worker:
-    image: langgenius/dify-api:0.15.0
+    image: langgenius/dify-api:0.14.2
     restart: always
     environment:
       CONSOLE_WEB_URL: ''
@@ -397,7 +397,7 @@ services:
 
   # Frontend web application.
   web:
-    image: langgenius/dify-web:0.15.0
+    image: langgenius/dify-web:0.14.2
     restart: always
     environment:
       # The base URL of console application api server, refers to the Console base URL of WEB service if console domain is

docker/.env.example

@@ -126,13 +126,10 @@ DIFY_PORT=5001
 # The number of API server workers, i.e., the number of workers.
 # Formula: number of cpu cores x 2 + 1 for sync, 1 for Gevent
 # Reference: https://docs.gunicorn.org/en/stable/design.html#how-many-workers
-SERVER_WORKER_AMOUNT=1
+SERVER_WORKER_AMOUNT=
 
 # Defaults to gevent. If using windows, it can be switched to sync or solo.
-SERVER_WORKER_CLASS=gevent
-
-# Default number of worker connections, the default is 10.
-SERVER_WORKER_CONNECTIONS=10
+SERVER_WORKER_CLASS=
 
 # Similar to SERVER_WORKER_CLASS.
 # If using windows, it can be switched to sync or solo.
@@ -929,5 +926,3 @@ CREATE_TIDB_SERVICE_JOB_ENABLED=false
 # Maximum number of submitted thread count in a ThreadPool for parallel node execution
 MAX_SUBMIT_COUNT=100
 
-# The maximum number of top-k value for RAG.
-TOP_K_MAX_VALUE=10

docker/docker-compose-template.yaml

@@ -2,7 +2,7 @@ x-shared-env: &shared-api-worker-env
 services:
   # API service
   api:
-    image: langgenius/dify-api:0.15.0
+    image: langgenius/dify-api:0.14.2
     restart: always
     environment:
       # Use the shared environment variables.
@@ -25,7 +25,7 @@ services:
   # worker service
   # The Celery worker for processing the queue.
   worker:
-    image: langgenius/dify-api:0.15.0
+    image: langgenius/dify-api:0.14.2
     restart: always
     environment:
       # Use the shared environment variables.
@@ -47,7 +47,7 @@ services:
 
   # Frontend web application.
   web:
-    image: langgenius/dify-web:0.15.0
+    image: langgenius/dify-web:0.14.2
     restart: always
     environment:
       CONSOLE_API_URL: ${CONSOLE_API_URL:-}

docker/docker-compose.yaml

@@ -32,9 +32,8 @@ x-shared-env: &shared-api-worker-env
   APP_MAX_EXECUTION_TIME: ${APP_MAX_EXECUTION_TIME:-1200}
   DIFY_BIND_ADDRESS: ${DIFY_BIND_ADDRESS:-0.0.0.0}
   DIFY_PORT: ${DIFY_PORT:-5001}
-  SERVER_WORKER_AMOUNT: ${SERVER_WORKER_AMOUNT:-1}
-  SERVER_WORKER_CLASS: ${SERVER_WORKER_CLASS:-gevent}
-  SERVER_WORKER_CONNECTIONS: ${SERVER_WORKER_CONNECTIONS:-10}
+  SERVER_WORKER_AMOUNT: ${SERVER_WORKER_AMOUNT:-}
+  SERVER_WORKER_CLASS: ${SERVER_WORKER_CLASS:-}
   CELERY_WORKER_CLASS: ${CELERY_WORKER_CLASS:-}
   GUNICORN_TIMEOUT: ${GUNICORN_TIMEOUT:-360}
   CELERY_WORKER_AMOUNT: ${CELERY_WORKER_AMOUNT:-}
@@ -387,12 +386,11 @@ x-shared-env: &shared-api-worker-env
   CSP_WHITELIST: ${CSP_WHITELIST:-}
   CREATE_TIDB_SERVICE_JOB_ENABLED: ${CREATE_TIDB_SERVICE_JOB_ENABLED:-false}
   MAX_SUBMIT_COUNT: ${MAX_SUBMIT_COUNT:-100}
-  TOP_K_MAX_VALUE: ${TOP_K_MAX_VALUE:-10}
 
 services:
   # API service
   api:
-    image: langgenius/dify-api:0.15.0
+    image: langgenius/dify-api:0.14.2
     restart: always
     environment:
       # Use the shared environment variables.
@@ -415,7 +413,7 @@ services:
   # worker service
   # The Celery worker for processing the queue.
   worker:
-    image: langgenius/dify-api:0.15.0
+    image: langgenius/dify-api:0.14.2
     restart: always
     environment:
       # Use the shared environment variables.
@@ -437,7 +435,7 @@ services:
 
   # Frontend web application.
   web:
-    image: langgenius/dify-web:0.15.0
+    image: langgenius/dify-web:0.14.2
     restart: always
     environment:
       CONSOLE_API_URL: ${CONSOLE_API_URL:-}

web/app/components/app/configuration/config-prompt/prompt-editor-height-resize-wrap.tsx

@@ -26,15 +26,13 @@ const PromptEditorHeightResizeWrap: FC<Props> = ({
   const [clientY, setClientY] = useState(0)
   const [isResizing, setIsResizing] = useState(false)
   const [prevUserSelectStyle, setPrevUserSelectStyle] = useState(getComputedStyle(document.body).userSelect)
-  const [oldHeight, setOldHeight] = useState(height)
 
   const handleStartResize = useCallback((e: React.MouseEvent<HTMLElement>) => {
     setClientY(e.clientY)
     setIsResizing(true)
-    setOldHeight(height)
     setPrevUserSelectStyle(getComputedStyle(document.body).userSelect)
     document.body.style.userSelect = 'none'
-  }, [height])
+  }, [])
 
   const handleStopResize = useCallback(() => {
     setIsResizing(false)
@@ -46,7 +44,8 @@ const PromptEditorHeightResizeWrap: FC<Props> = ({
       return
 
     const offset = e.clientY - clientY
-    let newHeight = oldHeight + offset
+    let newHeight = height + offset
+    setClientY(e.clientY)
     if (newHeight < minHeight)
       newHeight = minHeight
     onHeightChange(newHeight)

web/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dify-web",
-  "version": "0.15.0",
+  "version": "0.14.2",
   "private": true,
   "engines": {
     "node": ">=18.17.0"