truss 0.9.45.dev101__tar.gz → 0.9.45rc6__tar.gz

{truss-0.9.45.dev101 → truss-0.9.45rc6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: truss
-Version: 0.9.45.dev101
+Version: 0.9.45rc6
 Summary: A seamless bridge from model development to model delivery
 Home-page: https://github.com/basetenlabs/truss
 License: MIT
@@ -17,7 +17,6 @@ Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3.12
 Requires-Dist: Jinja2 (>=3.1.2,<4.0.0)
 Requires-Dist: PyYAML (>=6.0)
-Requires-Dist: aiohttp (>=3.10.10,<4.0.0)
 Requires-Dist: blake3 (>=0.3.3,<0.4.0)
 Requires-Dist: boto3 (>=1.34.85,<2.0.0)
 Requires-Dist: fastapi (>=0.109.1)

{truss-0.9.45.dev101 → truss-0.9.45rc6}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "truss"
-version = "0.9.45dev101"
+version = "0.9.45rc006"
 description = "A seamless bridge from model development to model delivery"
 license = "MIT"
 readme = "README.md"
@@ -27,7 +27,6 @@ packages = [
 "Baseten" = "https://baseten.co"
 
 [tool.poetry.dependencies]
-aiohttp = "^3.10.10"
 blake3 = "^0.3.3"
 boto3 = "^1.34.85"
 fastapi = ">=0.109.1"

{truss-0.9.45.dev101 → truss-0.9.45rc6}/truss/config/trt_llm.py

@@ -23,6 +23,7 @@ class TrussTRTLLMModel(str, Enum):
     MISTRAL = "mistral"
     DEEPSEEK = "deepseek"
     WHISPER = "whisper"
+    QWEN = "qwen"
 
 
 class TrussTRTLLMQuantizationType(str, Enum):
@@ -58,10 +59,9 @@ class CheckpointRepository(BaseModel):
 
 class TrussTRTLLMBuildConfiguration(BaseModel):
     base_model: TrussTRTLLMModel
-    max_input_len: int
-    max_output_len: int
-    max_batch_size: int
-    max_num_tokens: Optional[int] = None
+    max_seq_len: int
+    max_batch_size: Optional[int] = 256
+    max_num_tokens: Optional[int] = 8192
     max_beam_width: int = 1
     max_prompt_embedding_table_size: int = 0
     checkpoint_repository: CheckpointRepository

{truss-0.9.45.dev101 → truss-0.9.45rc6}/truss/test_data/test_trt_llm_truss/config.yaml

@@ -4,10 +4,9 @@ resources:
   use_gpu: True
 trt_llm:
   build:
-    max_input_len: 1000
+    max_seq_len: 1000
     max_batch_size: 1
     max_beam_width: 1
-    max_output_len: 1000
     base_model: llama
     checkpoint_repository:
       repo: TinyLlama/TinyLlama-1.1B-Chat-v1.0

{truss-0.9.45.dev101 → truss-0.9.45rc6}/truss/tests/conftest.py

@@ -389,8 +389,7 @@ def custom_model_trt_llm(tmp_path):
             content["trt_llm"] = {
                 "build": {
                     "base_model": "llama",
-                    "max_input_len": 1024,
-                    "max_output_len": 1024,
+                    "max_seq_len": 2048,
                     "max_batch_size": 512,
                     "checkpoint_repository": {
                         "source": "HF",

{truss-0.9.45.dev101 → truss-0.9.45rc6}/truss/tests/test_config.py

@@ -52,8 +52,7 @@ def trtllm_config(default_config) -> Dict[str, Any]:
     trtllm_config["trt_llm"] = {
         "build": {
             "base_model": "llama",
-            "max_input_len": 1024,
-            "max_output_len": 1024,
+            "max_seq_len": 2048,
             "max_batch_size": 512,
             "checkpoint_repository": {
                 "source": "HF",

{truss-0.9.45.dev101 → truss-0.9.45rc6}/truss-chains/truss_chains/stub.py

@@ -7,7 +7,6 @@ import time
 from typing import Any, ClassVar, Mapping, Optional, Type, TypeVar, final
 
 import httpx
-import aiohttp
 import tenacity
 
 from truss_chains import definitions, utils
@@ -82,20 +81,17 @@ class BasetenSession:
         assert self._cached_sync_client is not None
         return self._cached_sync_client[0]
 
-    async def _client_async(self) -> aiohttp.ClientSession:
+    async def _client_async(self) -> httpx.AsyncClient:
         # Check `_client_cycle_needed` before and after locking to avoid
         # needing a lock each time the client is accessed.
         if self._client_cycle_needed(self._cached_async_client):
             async with self._async_lock:
                 if self._client_cycle_needed(self._cached_async_client):
-                    connector = aiohttp.TCPConnector(
-                        limit=self._client_limits.max_connections,
-                    )
                     self._cached_async_client = (
-                        aiohttp.ClientSession(
+                        httpx.AsyncClient(
                             headers=self._auth_header,
-                            connector=connector,
-                            timeout=aiohttp.ClientTimeout(total=self._service_descriptor.options.timeout_sec),
+                            timeout=self._service_descriptor.options.timeout_sec,
+                            limits=self._client_limits,
                         ),
                         int(time.time()),
                     )
@@ -142,7 +138,7 @@ class BasetenSession:
                         resp = await client.post(
                             self._service_descriptor.predict_url, json=json_payload
                         )
-                    return await utils.handle_async_response(resp, self.name)
+                    return utils.handle_response(resp, self.name)
                 # As a special case we invalidate the client in case of certificate
                 # errors. This has happened in the past and is a defensive measure.
                 except ssl.SSLError:

{truss-0.9.45.dev101 → truss-0.9.45rc6}/truss-chains/truss_chains/utils.py

@@ -309,45 +309,6 @@ def handle_response(response: httpx.Response, remote_name: str) -> Any:
     return response.json()
 
 
-async def handle_async_response(response: httpx.Response, remote_name: str) -> Any:
-    """For successful requests returns JSON, otherwise raises error.
-    """
-    if response.status >= 400:
-        try:
-            response_json = await response.json()
-        except Exception as e:
-            raise ValueError(
-                "Could not get JSON from error response. Status: "
-                f"`{response.status}`."
-            ) from e
-        try:
-            error_json = response_json["error"]
-        except KeyError as e:
-            logging.error(f"response_json: {response_json}")
-            raise ValueError(
-                "Could not get `error` field from JSON from error response"
-            ) from e
-        try:
-            error = definitions.RemoteErrorDetail.model_validate(error_json)
-        except pydantic.ValidationError as e:
-            if isinstance(error_json, str):
-                msg = f"Remote error occurred in `{remote_name}`: '{error_json}'"
-                raise definitions.GenericRemoteException(msg) from None
-            raise ValueError(
-                "Could not parse error. Error details are expected to be either a "
-                "plain string (old truss models) or a serialized "
-                f"`definitions.RemoteErrorDetail.__name__`, got:\n{repr(error_json)}"
-            ) from e
-        exception_cls = _resolve_exception_class(error)
-        msg = (
-            f"(showing remote errors, root message at the bottom)\n"
-            f"--> Preceding Remote Cause:\n"
-            f"{textwrap.indent(error.format(), '    ')}"
-        )
-        raise exception_cls(msg)
-    return await response.json()
-
-
 class InjectedError(Exception):
     """Test error for debugging/dev."""
 

truss-0.9.45.dev101/truss/.pytest_cache/.gitignore

@@ -1,2 +0,0 @@
-# Created by pytest automatically.
-*

truss-0.9.45.dev101/truss/.pytest_cache/CACHEDIR.TAG

@@ -1,4 +0,0 @@
-Signature: 8a477f597d28d172789f06886806bc55
-# This file is a cache directory tag created by pytest.
-# For information about cache directory tags, see:
-#	https://bford.info/cachedir/spec.html

truss-0.9.45.dev101/truss/.pytest_cache/README.md

@@ -1,8 +0,0 @@
-# pytest cache directory #
-
-This directory contains data from the pytest's cache plugin,
-which provides the `--lf` and `--ff` options, as well as the `cache` fixture.
-
-**Do not** commit this to version control.
-
-See [the docs](https://docs.pytest.org/en/stable/how-to/cache.html) for more information.

truss-0.9.45.dev101/truss/.pytest_cache/v/cache/lastfailed

@@ -1 +0,0 @@
-{}

truss-0.9.45.dev101/truss/.pytest_cache/v/cache/nodeids

@@ -1,3 +0,0 @@
-[
-  "tests/test_control_truss_patching.py::test_patch_added_model_dir"
-]

truss-0.9.45.dev101/truss/.pytest_cache/v/cache/stepwise

@@ -1 +0,0 @@
-[]

truss-0.9.45.dev101/truss/test_data/test_basic_truss/packages2/package1/foo.py

@@ -1,2 +0,0 @@
-def bar():
-    print("in bar")

truss-0.9.45.dev101/truss/test_data/test_basic_truss/sidscript.py

@@ -1,8 +0,0 @@
-import os
-
-print("HI SID")
-
-print("Hi agaisdfn")
-print("more content")
-
-raise Exception("Foo")