guidellm 0.5.2__tar.gz → 0.5.4__tar.gz

{guidellm-0.5.2/src/guidellm.egg-info → guidellm-0.5.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: guidellm
-Version: 0.5.2
+Version: 0.5.4
 Summary: Guidance platform for deploying and managing large language models.
 Author: Red Hat
 License: Apache-2.0
@@ -33,17 +33,18 @@ Requires-Dist: transformers
 Requires-Dist: uvloop>=0.18
 Requires-Dist: torch
 Provides-Extra: all
-Requires-Dist: guidellm[audio,openai,perf,vision]; extra == "all"
+Requires-Dist: guidellm[audio,perf,tokenizers,vision]; extra == "all"
 Provides-Extra: recommended
-Requires-Dist: guidellm[openai,perf]; extra == "recommended"
+Requires-Dist: guidellm[perf,tokenizers]; extra == "recommended"
 Provides-Extra: perf
 Requires-Dist: orjson; extra == "perf"
 Requires-Dist: msgpack; extra == "perf"
 Requires-Dist: msgspec; extra == "perf"
 Requires-Dist: uvloop; extra == "perf"
-Provides-Extra: openai
-Requires-Dist: tiktoken>=0.11.0; extra == "openai"
-Requires-Dist: blobfile>=3.1.0; extra == "openai"
+Provides-Extra: tokenizers
+Requires-Dist: tiktoken; extra == "tokenizers"
+Requires-Dist: blobfile; extra == "tokenizers"
+Requires-Dist: mistral-common; extra == "tokenizers"
 Provides-Extra: audio
 Requires-Dist: datasets[audio]>=4.1.0; extra == "audio"
 Requires-Dist: torch==2.9.*; extra == "audio"

{guidellm-0.5.2 → guidellm-0.5.4}/pyproject.toml

@@ -74,11 +74,11 @@ dependencies = [
 
 [project.optional-dependencies]
 # Meta Extras
-all = ["guidellm[perf,openai,audio,vision]"]
-recommended = ["guidellm[perf,openai]"]
+all = ["guidellm[perf,tokenizers,audio,vision]"]
+recommended = ["guidellm[perf,tokenizers]"]
 # Feature Extras
 perf = ["orjson", "msgpack", "msgspec", "uvloop"]
-openai = ["tiktoken>=0.11.0", "blobfile>=3.1.0"]
+tokenizers = ["tiktoken", "blobfile", "mistral-common"]
 audio = [
     # Lowest version with full torchcodec support
     "datasets[audio]>=4.1.0",

{guidellm-0.5.2 → guidellm-0.5.4}/src/guidellm/__main__.py

@@ -334,7 +334,7 @@ def benchmark():
     default=BenchmarkGenerativeTextArgs.get_default("rampup"),
     help=(
         "The time, in seconds, to ramp up the request rate over. "
-        "Only applicable for Throughput/Concurrent strategies"
+        "Applicable for Throughput, Concurrent, and Constant strategies"
     ),
 )
 @click.option(

{guidellm-0.5.2 → guidellm-0.5.4}/src/guidellm/benchmark/benchmarker.py

@@ -64,7 +64,7 @@ class Benchmarker(
         environment: Environment,
         warmup: TransientPhaseConfig,
         cooldown: TransientPhaseConfig,
-        sample_requests: int | None = 20,
+        sample_requests: int | None = None,
         prefer_response_metrics: bool = True,
         progress: (
             BenchmarkerProgress[BenchmarkAccumulatorT, BenchmarkT] | None

{guidellm-0.5.2 → guidellm-0.5.4}/src/guidellm/benchmark/profiles.py

@@ -557,7 +557,9 @@ class AsyncProfile(Profile):
 
         if self.strategy_type == "constant":
             return AsyncConstantStrategy(
-                rate=current_rate, max_concurrency=self.max_concurrency
+                rate=current_rate,
+                max_concurrency=self.max_concurrency,
+                rampup_duration=self.rampup_duration,
             )
         elif self.strategy_type == "poisson":
             return AsyncPoissonStrategy(

{guidellm-0.5.2 → guidellm-0.5.4}/src/guidellm/benchmark/schemas/base.py

@@ -273,7 +273,7 @@ class BenchmarkConfig(StandardBaseDict):
         description="Constraint definitions applied to scheduler strategy execution",
     )
     sample_requests: int | None = Field(
-        default=20,
+        default=None,
         description="Request count for statistical sampling in final metrics",
     )
     warmup: TransientPhaseConfig = Field(

{guidellm-0.5.2 → guidellm-0.5.4}/src/guidellm/benchmark/schemas/generative/accumulator.py

@@ -787,6 +787,21 @@ class GenerativeBenchmarkAccumulator(
         description="Running metrics for incomplete requests",
     )
 
+    def model_post_init(self, __context):
+        """
+        Initialize child accumulators with config values after model construction.
+
+        Propagates sample_requests from config to child request accumulators to ensure
+        consistent sampling behavior across completed, errored, and incomplete request
+        collections. This ensures the --sample-requests option functions correctly.
+        """
+        super().model_post_init(__context)
+
+        # Propagate sample_requests from config to child accumulators
+        self.completed.sample_requests = self.config.sample_requests
+        self.errored.sample_requests = self.config.sample_requests
+        self.incomplete.sample_requests = self.config.sample_requests
+
     def update_estimate(
         self,
         response: GenerationResponse | None,

{guidellm-0.5.2 → guidellm-0.5.4}/src/guidellm/benchmark/schemas/generative/entrypoints.py

@@ -239,7 +239,7 @@ class BenchmarkGenerativeTextArgs(StandardBaseModel):
     )
     # Benchmarker configuration
     sample_requests: int | None = Field(
-        default=10,
+        default=None,
         description="Number of requests to sample for detailed metrics (None for all)",
     )
     warmup: int | float | dict | TransientPhaseConfig | None = Field(

{guidellm-0.5.2 → guidellm-0.5.4}/src/guidellm/data/deserializers/synthetic.py

@@ -150,7 +150,7 @@ class _SyntheticTextExamplesIterable(_BaseExamplesIterable):
             text = unique + faker.text(max_nb_chars=num_chars)
             prompt_token_ids = self.processor.encode(text)
 
-        return self.processor.decode(
+        return self.processor.decode(  # type: ignore[return-value]
             prompt_token_ids[:prompt_tokens_count], skip_special_tokens=True
         )
 

{guidellm-0.5.2 → guidellm-0.5.4}/src/guidellm/data/processor.py

@@ -3,7 +3,7 @@ from __future__ import annotations
 from pathlib import Path
 from typing import Any
 
-from transformers import AutoTokenizer, PreTrainedTokenizerBase  # type: ignore[import]
+from transformers import AutoTokenizer, PreTrainedTokenizerBase
 
 __all__ = ["ProcessorFactory"]
 

{guidellm-0.5.2 → guidellm-0.5.4}/src/guidellm/mock_server/handlers/tokenizer.py

@@ -13,7 +13,7 @@ from pydantic import ValidationError
 from sanic import response
 from sanic.request import Request
 from sanic.response import HTTPResponse
-from transformers.tokenization_utils import PreTrainedTokenizer
+from transformers import AutoTokenizer
 
 from guidellm.mock_server.config import MockServerConfig
 from guidellm.mock_server.models import (
@@ -55,7 +55,7 @@ class TokenizerHandler:
         self.tokenizer = (
             MockTokenizer()
             if config.processor is None
-            else PreTrainedTokenizer.from_pretrained(config.processor)
+            else AutoTokenizer.from_pretrained(config.processor)
         )
 
     async def tokenize(self, request: Request) -> HTTPResponse:
@@ -96,6 +96,9 @@ class TokenizerHandler:
 
         tokens = self.tokenizer.tokenize(req_data.text)
         token_ids = self.tokenizer.convert_tokens_to_ids(tokens)
+        # This should never happen in our mock tokenizer, but matches function sig
+        if not isinstance(token_ids, list):
+            token_ids = [token_ids]
 
         return response.json(
             TokenizeResponse(tokens=token_ids, count=len(token_ids)).model_dump()
@@ -139,4 +142,4 @@ class TokenizerHandler:
 
         text = self.tokenizer.decode(req_data.tokens, skip_special_tokens=False)
 
-        return response.json(DetokenizeResponse(text=text).model_dump())
+        return response.json(DetokenizeResponse(text=text).model_dump())  # type: ignore[arg-type]

{guidellm-0.5.2 → guidellm-0.5.4}/src/guidellm/mock_server/utils.py

@@ -14,7 +14,11 @@ import re
 from collections.abc import Generator
 
 from faker import Faker
-from transformers.tokenization_utils import AddedToken, PreTrainedTokenizer, TextInput
+from transformers.tokenization_utils_base import (
+    AddedToken,
+    PreTrainedTokenizerBase,
+    TextInput,
+)
 
 __all__ = [
     "MockTokenizer",
@@ -25,7 +29,7 @@ __all__ = [
 ]
 
 
-class MockTokenizer(PreTrainedTokenizer):
+class MockTokenizer(PreTrainedTokenizerBase):
     """
     Mock tokenizer implementation for testing text processing workflows.
 
@@ -207,7 +211,7 @@ class MockTokenizer(PreTrainedTokenizer):
 
 def create_fake_text(
     num_tokens: int,
-    processor: PreTrainedTokenizer,
+    processor: PreTrainedTokenizerBase,
     seed: int = 42,
     fake: Faker | None = None,
 ) -> str:
@@ -229,7 +233,7 @@ def create_fake_text(
 
 def create_fake_tokens_str(
     num_tokens: int,
-    processor: PreTrainedTokenizer,
+    processor: PreTrainedTokenizerBase,
     seed: int = 42,
     fake: Faker | None = None,
 ) -> list[str]:

{guidellm-0.5.2 → guidellm-0.5.4}/src/guidellm/scheduler/strategies.py

@@ -16,6 +16,7 @@ throughput (maximum load), constant-rate (steady intervals), and Poisson-distrib
 from __future__ import annotations
 
 import asyncio
+import math
 import random
 from abc import abstractmethod
 from multiprocessing import Event, Value, synchronize
@@ -453,6 +454,13 @@ class AsyncConstantStrategy(SchedulingStrategy):
         default=None,
         description="Maximum number of concurrent requests to schedule",
     )
+    rampup_duration: NonNegativeFloat = Field(
+        default=0.0,
+        description=(
+            "Duration in seconds to linearly ramp up from 0 to target rate "
+            "at the beginning of each strategy run"
+        ),
+    )
 
     def __str__(self) -> str:
         """
@@ -476,19 +484,47 @@ class AsyncConstantStrategy(SchedulingStrategy):
 
     async def next_request_time(self, worker_index: PositiveInt) -> float:
         """
-        Calculate next request time at fixed intervals.
+        Calculate next request time at fixed intervals with optional linear rampup.
 
         Schedules requests at uniform intervals determined by the configured rate,
-        independent of request completion times.
+        independent of request completion times. If rampup_duration is set, the rate
+        increases linearly from 0 to the target rate during the rampup period, then
+        continues at the constant rate.
 
         :param worker_index: Unused for constant strategy
-        :return: Start time plus constant interval based on request index
+        :return: Start time plus interval based on request index and
+            rampup configuration
         """
         _ = worker_index  # unused
         current_index = self.next_request_index()
         start_time = await self.get_processes_start_time()
 
-        return start_time + current_index / self.rate
+        if self.rampup_duration > 0:
+            # Calculate number of requests that would be sent during rampup
+            # Cumulative requests by time t during rampup:
+            # n = rate * t² / (2 * rampup_duration)
+            # At end of rampup (t = rampup_duration), n_rampup is calculated below
+            n_rampup = self.rate * self.rampup_duration / 2.0
+
+            if current_index == 1:
+                # First request at start_time
+                return start_time
+            elif current_index <= n_rampup:
+                # During rampup: solve for t where
+                # n = rate * t² / (2 * rampup_duration)
+                time_offset = math.sqrt(
+                    2.0 * current_index * self.rampup_duration / self.rate
+                )
+                return start_time + time_offset
+            else:
+                # After rampup: continue at constant rate
+                time_offset = (
+                    self.rampup_duration + (current_index - n_rampup) / self.rate
+                )
+                return start_time + time_offset
+        else:
+            # No rampup: uniform intervals
+            return start_time + current_index / self.rate
 
     def request_completed(self, request_info: RequestInfo):
         """

{guidellm-0.5.2 → guidellm-0.5.4}/src/guidellm/settings.py

@@ -31,9 +31,9 @@ class Environment(str, Enum):
 
 
 ENV_REPORT_MAPPING = {
-    Environment.PROD: "https://blog.vllm.ai/guidellm/ui/v0.5.0/index.html",
-    Environment.STAGING: "https://blog.vllm.ai/guidellm/ui/release/v0.4.0/index.html",
-    Environment.DEV: "https://blog.vllm.ai/guidellm/ui/dev/index.html",
+    Environment.PROD: "https://raw.githubusercontent.com/vllm-project/guidellm/refs/heads/gh-pages/ui/v0.5.3/index.html",
+    Environment.STAGING: "https://raw.githubusercontent.com/vllm-project/guidellm/refs/heads/gh-pages/ui/release/v0.4.0/index.html",
+    Environment.DEV: "https://raw.githubusercontent.com/vllm-project/guidellm/refs/heads/gh-pages/ui/dev/index.html",
     Environment.LOCAL: "http://localhost:3000/index.html",
 }
 

{guidellm-0.5.2 → guidellm-0.5.4}/src/guidellm/utils/hf_transformers.py

@@ -1,7 +1,8 @@
 from pathlib import Path
 from typing import Any
 
-from transformers import AutoTokenizer, PreTrainedTokenizerBase  # type: ignore[import]
+from transformers import AutoTokenizer
+from transformers.tokenization_utils_base import PreTrainedTokenizerBase
 
 __all__ = [
     "check_load_processor",
@@ -16,6 +17,7 @@ def check_load_processor(
     if processor is None:
         raise ValueError(f"Processor/Tokenizer is required for {error_msg}.")
 
+    loaded: PreTrainedTokenizerBase
     try:
         if isinstance(processor, str | Path):
             loaded = AutoTokenizer.from_pretrained(

guidellm-0.5.4/src/guidellm/version.py

@@ -0,0 +1,6 @@
+version = "0.5.4"
+build_type = "release"
+build_iteration = "0"
+git_commit = "fc31a9ebfaef291d141cb1fbfd64ad82dec9ecbf"
+git_branch = "HEAD"
+git_last_tag = "v0.5.4"

{guidellm-0.5.2 → guidellm-0.5.4/src/guidellm.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: guidellm
-Version: 0.5.2
+Version: 0.5.4
 Summary: Guidance platform for deploying and managing large language models.
 Author: Red Hat
 License: Apache-2.0
@@ -33,17 +33,18 @@ Requires-Dist: transformers
 Requires-Dist: uvloop>=0.18
 Requires-Dist: torch
 Provides-Extra: all
-Requires-Dist: guidellm[audio,openai,perf,vision]; extra == "all"
+Requires-Dist: guidellm[audio,perf,tokenizers,vision]; extra == "all"
 Provides-Extra: recommended
-Requires-Dist: guidellm[openai,perf]; extra == "recommended"
+Requires-Dist: guidellm[perf,tokenizers]; extra == "recommended"
 Provides-Extra: perf
 Requires-Dist: orjson; extra == "perf"
 Requires-Dist: msgpack; extra == "perf"
 Requires-Dist: msgspec; extra == "perf"
 Requires-Dist: uvloop; extra == "perf"
-Provides-Extra: openai
-Requires-Dist: tiktoken>=0.11.0; extra == "openai"
-Requires-Dist: blobfile>=3.1.0; extra == "openai"
+Provides-Extra: tokenizers
+Requires-Dist: tiktoken; extra == "tokenizers"
+Requires-Dist: blobfile; extra == "tokenizers"
+Requires-Dist: mistral-common; extra == "tokenizers"
 Provides-Extra: audio
 Requires-Dist: datasets[audio]>=4.1.0; extra == "audio"
 Requires-Dist: torch==2.9.*; extra == "audio"

{guidellm-0.5.2 → guidellm-0.5.4}/src/guidellm.egg-info/requires.txt

@@ -20,7 +20,7 @@ uvloop>=0.18
 torch
 
 [all]
-guidellm[audio,openai,perf,vision]
+guidellm[audio,perf,tokenizers,vision]
 
 [audio]
 datasets[audio]>=4.1.0
@@ -56,10 +56,6 @@ types-requests~=2.32.0
 types-toml
 mkdocs-linkcheck~=1.0.6
 
-[openai]
-tiktoken>=0.11.0
-blobfile>=3.1.0
-
 [perf]
 orjson
 msgpack
@@ -67,7 +63,12 @@ msgspec
 uvloop
 
 [recommended]
-guidellm[openai,perf]
+guidellm[perf,tokenizers]
+
+[tokenizers]
+tiktoken
+blobfile
+mistral-common
 
 [vision]
 datasets[vision]

guidellm-0.5.2/src/guidellm/version.py

@@ -1,6 +0,0 @@
-version = "0.5.2"
-build_type = "release"
-build_iteration = "0"
-git_commit = "26dfc10677055a08fdaac470839983222f1ee7ff"
-git_branch = "HEAD"
-git_last_tag = "v0.5.2"