Upload 6 files

app.py

@@ -204,6 +204,56 @@ def _recompute_out_and_state_from_tokens(model_name: str, model_tokens: List[int
         tokens = tokens[CONFIG.CHUNK_LEN :]
     return out, model_state
 
+
+
+# Move ChatCompletionRequest definition above fallback apply_model_tags_to_request
+
+
+# Move ChatCompletionRequest definition above fallback apply_model_tags_to_request
+
+try:
+    from model_tags import apply_model_tags_to_request_obj as apply_model_tags_to_request
+except Exception:
+    def apply_model_tags_to_request(req: Any):
+        # Fallback implementation if the module import fails; keep behavior robust
+        if not req or not getattr(req, 'model', None) or ':' not in req.model:
+            return
+        original = req.model
+        parts = [p.strip() for p in original.split(":") if p is not None and p != ""]
+        if len(parts) <= 1:
+            return
+        base = parts[0]
+        tags = parts[1:]
+        req.model = base
+        for tag in tags:
+            t = tag.lower()
+            if t in ("thinking", "think", "reasoning", "reason"):
+                req.enable_reasoning = True
+                req.auto_reasoning = False
+            elif t in ("web", "web_search", "search"):
+                req.enable_web_search = True
+                req.web_search = True
+                req.auto_web_search = False
+            elif t in ("no-web", "disable-web", "no-web-search"):
+                req.enable_web_search = False
+                req.web_search = False
+            elif t in ("tools", "enable-tools"):
+                req.enable_tools = True
+                req.auto_tools = False
+            elif t in ("no-tools", "disable-tools"):
+                req.enable_tools = False
+            elif t in ("file", "file_tool", "filetool"):
+                req.enable_file_tool = True
+                req.auto_file_tool = False
+            elif t in ("no-file", "disable-file"):
+                req.enable_file_tool = False
+            elif t in ("universal", "univ"):
+                req.enable_universal = True
+                req.auto_universal = False
+            elif t in ("stream",):
+                req.stream = True
+
+
 logger.info(f"STRATEGY - {CONFIG.STRATEGY}")
 
 logGPUState()
@@ -1604,9 +1654,11 @@ async def chat_completions(request: ChatCompletionRequest):
     completionId = str(next(CompletionIdGenerator))
     logger.info(f"[REQ] {completionId} - {request.model_dump()}")
 
-    # Model suffix tags are not allowed. Instruct clients to use request flags instead.
-    if ":" in request.model:
-        raise HTTPException(400, "Model tags/suffixes (e.g., ':thinking', ':web') are deprecated. Please set enable_reasoning, web_search, and enable_file_tool in the request body instead.")
+    # Apply any legacy model suffix tags (e.g., 'rwkv-latest:thinking' -> enable_reasoning)
+    # This helper is defined at module level so it can be unit-tested and reused.
+
+    # Apply legacy tags (if present) to request and proceed normally
+    apply_model_tags_to_request(request)
     modelName = request.model
 
     if request.model == "rwkv-latest":

config.local.yaml

@@ -1,8 +1,13 @@
 HOST: "0.0.0.0"
 PORT: 7860
-STRATEGY: "cpu fp16"
-RWKV_CUDA_ON: False
+STRATEGY: "cpu fp16"
+RWKV_CUDA_ON: False
 CHUNK_LEN: 256
+DEFAULT_STREAM: True
+AUTO_ENABLE_TOOLS: True
+AUTO_ENABLE_REASONING: True
+AUTO_ENABLE_WEB_SEARCH: True
+ENABLE_TOOLS_BY_DEFAULT: False
 MODELS:
   - SERVICE_NAME: "rwkv7-g1a-0.1b-20250728-ctx4096"
     DOWNLOAD_MODEL_FILE_NAME: "rwkv7-g1a-0.1b-20250728-ctx4096.pth"
@@ -11,6 +16,9 @@ MODELS:
     REASONING: True
     DEFAULT_CHAT: True
     DEFAULT_REASONING: True
+    ALLOW_WEB_SEARCH: True
+    ALLOW_TOOLS: True
+    ALLOW_REASONING: True
     DEFAULT_SAMPLER:
       max_tokens: 4096
       temperature: 1.0
@@ -25,3 +33,6 @@ MODELS:
       ALLOW_WEB_SEARCH: True
       ALLOW_TOOLS: True
       ALLOW_REASONING: True
+STATE_STORE_PATH: "./state_store.json"
+STATE_STORE_FLUSH_INTERVAL: 5
+STATE_STORE_SAVE_ON_UPDATE: True

model_tags.py

@@ -0,0 +1,90 @@
+"""Utility module for parsing model tags like 'model:thinking:web' and
+applying them as request flags on a request-like object.
+
+The functions here operate on any object with the expected attributes (model,
+enable_reasoning, auto_reasoning, enable_web_search, web_search, auto_web_search,
+enable_tools, auto_tools, enable_file_tool, auto_file_tool, enable_universal,
+auto_universal, stream) and don't require heavy app imports, which makes them
+easy to unit-test.
+"""
+from types import SimpleNamespace
+from typing import Any
+
+
+def apply_model_tags_to_request_obj(req: Any) -> None:
+    """Apply model suffix tags to a request-like object.
+
+    The `req` can be a SimpleNamespace, an object, or a pydantic model with
+    mutable attributes. The function mutates `req.model` to the base model and
+    sets boolean flags for detected tags.
+    """
+    if not req or not getattr(req, 'model', None) or ':' not in req.model:
+        return
+    original = req.model
+    parts = [p.strip() for p in original.split(":") if p is not None and p != ""]
+    if len(parts) <= 1:
+        return
+    base = parts[0]
+    tags = parts[1:]
+    try:
+        req.model = base
+    except Exception:
+        # If object is read-only for model, don't fail
+        pass
+    for tag in tags:
+        t = tag.lower()
+        if t in ("thinking", "think", "reasoning", "reason"):
+            try:
+                setattr(req, 'enable_reasoning', True)
+                setattr(req, 'auto_reasoning', False)
+            except Exception:
+                pass
+        elif t in ("web", "web_search", "search"):
+            try:
+                setattr(req, 'enable_web_search', True)
+                setattr(req, 'web_search', True)
+                setattr(req, 'auto_web_search', False)
+            except Exception:
+                pass
+        elif t in ("no-web", "disable-web", "no-web-search"):
+            try:
+                setattr(req, 'enable_web_search', False)
+                setattr(req, 'web_search', False)
+            except Exception:
+                pass
+        elif t in ("tools", "enable-tools"):
+            try:
+                setattr(req, 'enable_tools', True)
+                setattr(req, 'auto_tools', False)
+            except Exception:
+                pass
+        elif t in ("no-tools", "disable-tools"):
+            try:
+                setattr(req, 'enable_tools', False)
+            except Exception:
+                pass
+        elif t in ("file", "file_tool", "filetool"):
+            try:
+                setattr(req, 'enable_file_tool', True)
+                setattr(req, 'auto_file_tool', False)
+            except Exception:
+                pass
+        elif t in ("no-file", "disable-file"):
+            try:
+                setattr(req, 'enable_file_tool', False)
+            except Exception:
+                pass
+        elif t in ("universal", "univ"):
+            try:
+                setattr(req, 'enable_universal', True)
+                setattr(req, 'auto_universal', False)
+            except Exception:
+                pass
+        elif t in ("stream",):
+            try:
+                setattr(req, 'stream', True)
+            except Exception:
+                pass
+        else:
+            # Unknown tags are ignored; callers may log if needed
+            pass