truss 0.11.8rc7__py3-none-any.whl → 0.11.9rc1__py3-none-any.whl

truss/base/trt_llm_config.py

@@ -329,9 +329,16 @@ pip install truss==0.10.8
             raise ValueError("Using fp8 context fmha requires paged context fmha")
         if (
             self.plugin_configuration.use_fp8_context_fmha
-            and not self.quantization_type == TrussTRTLLMQuantizationType.FP8_KV
+            and self.quantization_type
+            not in (
+                TrussTRTLLMQuantizationType.FP8_KV,
+                TrussTRTLLMQuantizationType.FP4_KV,
+            )
         ):
-            raise ValueError("Using fp8 context fmha requires fp8 kv cache dtype")
+            raise ValueError(
+                "Using fp8 context fmha requires fp8 kv, or fp4 with kv cache dtype"
+            )
+
         return self
 
     def _validate_speculator_config(self):

truss/templates/control/control/application.py

@@ -2,13 +2,12 @@ import asyncio
 import logging
 import logging.config
 import re
-import traceback
 from pathlib import Path
-from typing import Awaitable, Callable, Dict
+from typing import Dict
 
 import httpx
 from endpoints import control_app
-from fastapi import FastAPI, Request, Response
+from fastapi import FastAPI
 from fastapi.responses import JSONResponse
 from helpers.errors import ModelLoadFailed, PatchApplicatonError
 from helpers.inference_server_controller import InferenceServerController
@@ -17,45 +16,22 @@ from helpers.inference_server_starter import async_inference_server_startup_flow
 from helpers.truss_patch.model_container_patch_applier import ModelContainerPatchApplier
 from shared import log_config
 from starlette.datastructures import State
-from starlette.middleware.base import BaseHTTPMiddleware
-
-SANITIZED_EXCEPTION_FRAMES = 2
-
-
-class SanitizedExceptionMiddleware(BaseHTTPMiddleware):
-    def __init__(self, app, num_frames: int = SANITIZED_EXCEPTION_FRAMES):
-        super().__init__(app)
-        self.num_frames = num_frames
-
-    async def dispatch(
-        self, request: Request, call_next: Callable[[Request], Awaitable[Response]]
-    ) -> Response:
-        try:
-            return await call_next(request)
-        except Exception as exc:
-            sanitized_traceback = self._create_sanitized_traceback(exc)
-            if hasattr(request.app.state, "logger"):
-                request.app.state.logger.error(f"Error:\n{sanitized_traceback}")
-
-            if isinstance(exc, ModelLoadFailed):
-                return JSONResponse({"error": str(exc)}, status_code=503)
-            elif isinstance(exc, PatchApplicatonError):
-                error_type = _camel_to_snake_case(type(exc).__name__)
-                return JSONResponse(
-                    {"error": {"type": error_type, "msg": str(exc)}}, status_code=400
-                )
-            else:
-                return JSONResponse(
-                    {"error": {"type": "unknown", "msg": str(exc)}}, status_code=500
-                )
-
-    def _create_sanitized_traceback(self, error: Exception) -> str:
-        tb_lines = traceback.format_tb(error.__traceback__)
-        if tb_lines and self.num_frames > 0:
-            selected_frames = tb_lines[-self.num_frames :]
-            return "".join(selected_frames).rstrip()
-        else:
-            return f"{type(error).__name__}: {error}"
+
+
+async def handle_patch_error(_, exc):
+    error_type = _camel_to_snake_case(type(exc).__name__)
+    return JSONResponse(content={"error": {"type": error_type, "msg": str(exc)}})
+
+
+async def generic_error_handler(_, exc):
+    return JSONResponse(
+        content={"error": {"type": "unknown", "msg": f"{type(exc)}: {exc}"}}
+    )
+
+
+async def handle_model_load_failed(_, error):
+    # Model load failures should result in 503 status
+    return JSONResponse({"error": str(error)}, 503)
 
 
 def create_app(base_config: Dict):
@@ -106,10 +82,14 @@ def create_app(base_config: Dict):
     app = FastAPI(
         title="Truss Live Reload Server",
         on_startup=[start_background_inference_startup],
+        exception_handlers={
+            PatchApplicatonError: handle_patch_error,
+            ModelLoadFailed: handle_model_load_failed,
+            Exception: generic_error_handler,
+        },
     )
     app.state = app_state
     app.include_router(control_app)
-    app.add_middleware(SanitizedExceptionMiddleware)
 
     @app.on_event("shutdown")
     def on_shutdown():

truss/templates/control/control/endpoints.py

@@ -5,7 +5,6 @@ from typing import Any, Callable, Dict, Optional, Protocol
 import httpx
 from fastapi import APIRouter, WebSocket
 from fastapi.responses import JSONResponse, StreamingResponse
-from helpers.errors import ModelLoadFailed, ModelNotReady
 from httpx_ws import AsyncWebSocketSession, WebSocketDisconnect, aconnect_ws
 from httpx_ws import _exceptions as httpx_ws_exceptions
 from starlette.requests import ClientDisconnect, Request
@@ -14,6 +13,11 @@ from starlette.websockets import WebSocketDisconnect as StartletteWebSocketDisco
 from tenacity import RetryCallState, Retrying, retry_if_exception_type, wait_fixed
 from wsproto.events import BytesMessage, TextMessage
 
+from truss.templates.control.control.helpers.errors import (
+    ModelLoadFailed,
+    ModelNotReady,
+)
+
 INFERENCE_SERVER_START_WAIT_SECS = 60
 BASE_RETRY_EXCEPTIONS = (
     retry_if_exception_type(httpx.ConnectError)

{truss-0.11.8rc7.dist-info → truss-0.11.9rc1.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: truss
-Version: 0.11.8rc7
+Version: 0.11.9rc1
 Summary: A seamless bridge from model development to model delivery
 Project-URL: Repository, https://github.com/basetenlabs/truss
 Project-URL: Homepage, https://truss.baseten.co

{truss-0.11.8rc7.dist-info → truss-0.11.9rc1.dist-info}/RECORD

@@ -5,7 +5,7 @@ truss/base/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 truss/base/constants.py,sha256=sExArdnuGg83z83XMgaQ4b8SS3V_j_bJEpOATDGJzpE,3600
 truss/base/custom_types.py,sha256=FUSIT2lPOQb6gfg6IzT63YBV8r8L6NIZ0D74Fp3e_jQ,2835
 truss/base/errors.py,sha256=zDVLEvseTChdPP0oNhBBQCtQUtZJUaof5zeWMIjqz6o,691
-truss/base/trt_llm_config.py,sha256=CRz3AqGDAyv8YpcBWXUrnfjvNAauyo3yf8ZOGVsSt6g,32782
+truss/base/trt_llm_config.py,sha256=-hRpRsbxnfDaKS-5112yT0iP6R0evOtoTvnn557cwvc,32926
 truss/base/truss_config.py,sha256=7CtiJIwMHtDU8Wzn8UTJUVVunD0pWFl4QUVycK2aIpY,28055
 truss/base/truss_spec.py,sha256=jFVF79CXoEEspl2kXBAPyi-rwISReIGTdobGpaIhwJw,5979
 truss/cli/chains_commands.py,sha256=Kpa5mCg6URAJQE2ZmZfVQFhjBHEitKT28tKiW0H6XAI,17406
@@ -73,8 +73,8 @@ truss/templates/copy_cache_files.Dockerfile.jinja,sha256=Os5zFdYLZ_AfCRGq4RcpVTO
 truss/templates/docker_server_requirements.txt,sha256=PyhOPKAmKW1N2vLvTfLMwsEtuGpoRrbWuNo7tT6v2Mc,18
 truss/templates/server.Dockerfile.jinja,sha256=CUYnF_hgxPGq2re7__0UPWlwzOHMoFkxp6NVKi3U16s,7071
 truss/templates/control/requirements.txt,sha256=nqqNmlTwFeV8sV4fqwItwzzd_egADBP_e-cEopXBJ4k,358
-truss/templates/control/control/application.py,sha256=XfA5udraulB0z6s4J_S05w0y1TJGcoMuC3jJOjbPVu4,4839
-truss/templates/control/control/endpoints.py,sha256=KzqsLVNJE6r6TCPW8D5FMCtsfHadTwR15A3z_viGxmM,11782
+truss/templates/control/control/application.py,sha256=jYeta6hWe1SkfLL3W4IDmdYjg3ZuKqI_UagWYs5RB_E,3793
+truss/templates/control/control/endpoints.py,sha256=VQ1lvZjFvR091yRkiFdvXw1Q7PiNGXT9rJwY7_sX6yg,11828
 truss/templates/control/control/server.py,sha256=R4Y219i1dcz0kkksN8obLoX-YXWGo9iW1igindyG50c,3128
 truss/templates/control/control/helpers/context_managers.py,sha256=W6dyFgLBhPa5meqrOb3w_phMtKfaJI-GhwUfpiycDc8,413
 truss/templates/control/control/helpers/custom_types.py,sha256=n_lTudtLTpy4oPV3aDdJ4X2rh3KCV5btYO9UnTeUouQ,5471
@@ -368,8 +368,8 @@ truss_train/deployment.py,sha256=lWWANSuzBWu2M4oK4qD7n-oVR1JKdmw2Pn5BJQHg-Ck,307
 truss_train/loader.py,sha256=0o66EjBaHc2YY4syxxHVR4ordJWs13lNXnKjKq2wq0U,1630
 truss_train/public_api.py,sha256=9N_NstiUlmBuLUwH_fNG_1x7OhGCytZLNvqKXBlStrM,1220
 truss_train/restore_from_checkpoint.py,sha256=8hdPm-WSgkt74HDPjvCjZMBpvA9MwtoYsxVjOoa7BaM,1176
-truss-0.11.8rc7.dist-info/METADATA,sha256=37bkH41Vl6H2ftd6R4aJZWnxqanKAupqM4uOgFNmYlA,6680
-truss-0.11.8rc7.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-truss-0.11.8rc7.dist-info/entry_points.txt,sha256=-MwKfHHQHQ6j0HqIgvxrz3CehCmczDLTD-OsRHnjjuU,130
-truss-0.11.8rc7.dist-info/licenses/LICENSE,sha256=FTqGzu85i-uw1Gi8E_o0oD60bH9yQ_XIGtZbA1QUYiw,1064
-truss-0.11.8rc7.dist-info/RECORD,,
+truss-0.11.9rc1.dist-info/METADATA,sha256=nvg4yIXu46U0vuOQcptDUP_17YL3LGeFmsFmLPMgZbY,6680
+truss-0.11.9rc1.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+truss-0.11.9rc1.dist-info/entry_points.txt,sha256=-MwKfHHQHQ6j0HqIgvxrz3CehCmczDLTD-OsRHnjjuU,130
+truss-0.11.9rc1.dist-info/licenses/LICENSE,sha256=FTqGzu85i-uw1Gi8E_o0oD60bH9yQ_XIGtZbA1QUYiw,1064
+truss-0.11.9rc1.dist-info/RECORD,,