llama-benchy 0.1.0__tar.gz → 0.1.1__tar.gz

{llama_benchy-0.1.0 → llama_benchy-0.1.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: llama-benchy
-Version: 0.1.0
+Version: 0.1.1
 Summary: llama-bench style benchmarking tool for all OpenAI-compatible LLM endpoints
 Author: eugr
 License: MIT License
@@ -74,12 +74,26 @@ As of January 2nd, 2026, I wasn't able to find any existing benchmarking tool th
 - Supports executing a command after each run (e.g., to clear cache).
 - Configurable latency measurement mode.
 
+# Current Limitations
+
+- Evaluates against `/v1/chat/completions` endpoint only.
+- Doesn't measure throughput in concurrency mode (coming later).
+- Outputs results as a Markdown table only for now.
+
 ## Installation
 
-Install `uv` first: https://docs.astral.sh/uv/getting-started/installation/
+Using `uv` is recommended. You can install `uv` here: https://docs.astral.sh/uv/getting-started/installation/
 
 ### Option 1: Run without installation using `uvx`
 
+Run the release version from PyPI:
+
+```bash
+uvx llama-benchy --base-url <ENDPOINT_URL> --model <MODEL_NAME>
+```
+
+Run the latest version from the main branch:
+
 ```bash
 uvx --from git+https://github.com/eugr/llama-benchy llama-benchy --base-url <ENDPOINT_URL> --model <MODEL_NAME>
 ```
@@ -124,6 +138,14 @@ uv run llama-benchy --base-url <ENDPOINT_URL> --model <MODEL_NAME>
 
 ### Option 3: Install into system path
 
+Release version from PyPI:
+
+```bash
+uv pip install -U llama-benchy
+```
+
+Current version from the main branch:
+
 ```bash
 uv pip install git+https://github.com/eugr/llama-benchy --system
 ```
@@ -231,7 +253,7 @@ When `--enable-prefix-caching` is used (with `--depth` > 0), the script performs
 2.  **Inference**: Sends the same context (system message) followed by the actual prompt (user message). The server should reuse the cached context.
     -   Reported as standard `pp{tokens} @ d{depth}` and `tg{tokens} @ d{depth}`.
 
-By comparing the `pp` speed of the Inference step with a non-cached run (or the Context Load step), you can observe the speedup provided by prefix caching.
+In this case, `pp` and `tg` speeds will show an actual prompt processing / token generation speeds for a follow up prompt with a context pre-filled.
 
 ### Example
 

{llama_benchy-0.1.0 → llama_benchy-0.1.1}/README.md

@@ -29,12 +29,26 @@ As of January 2nd, 2026, I wasn't able to find any existing benchmarking tool th
 - Supports executing a command after each run (e.g., to clear cache).
 - Configurable latency measurement mode.
 
+# Current Limitations
+
+- Evaluates against `/v1/chat/completions` endpoint only.
+- Doesn't measure throughput in concurrency mode (coming later).
+- Outputs results as a Markdown table only for now.
+
 ## Installation
 
-Install `uv` first: https://docs.astral.sh/uv/getting-started/installation/
+Using `uv` is recommended. You can install `uv` here: https://docs.astral.sh/uv/getting-started/installation/
 
 ### Option 1: Run without installation using `uvx`
 
+Run the release version from PyPI:
+
+```bash
+uvx llama-benchy --base-url <ENDPOINT_URL> --model <MODEL_NAME>
+```
+
+Run the latest version from the main branch:
+
 ```bash
 uvx --from git+https://github.com/eugr/llama-benchy llama-benchy --base-url <ENDPOINT_URL> --model <MODEL_NAME>
 ```
@@ -79,6 +93,14 @@ uv run llama-benchy --base-url <ENDPOINT_URL> --model <MODEL_NAME>
 
 ### Option 3: Install into system path
 
+Release version from PyPI:
+
+```bash
+uv pip install -U llama-benchy
+```
+
+Current version from the main branch:
+
 ```bash
 uv pip install git+https://github.com/eugr/llama-benchy --system
 ```
@@ -186,7 +208,7 @@ When `--enable-prefix-caching` is used (with `--depth` > 0), the script performs
 2.  **Inference**: Sends the same context (system message) followed by the actual prompt (user message). The server should reuse the cached context.
     -   Reported as standard `pp{tokens} @ d{depth}` and `tg{tokens} @ d{depth}`.
 
-By comparing the `pp` speed of the Inference step with a non-cached run (or the Context Load step), you can observe the speedup provided by prefix caching.
+In this case, `pp` and `tg` speeds will show an actual prompt processing / token generation speeds for a follow up prompt with a context pre-filled.
 
 ### Example
 

llama_benchy-0.1.1/src/llama_benchy/__init__.py

@@ -0,0 +1,8 @@
+"""
+llama-benchy - llama-bench style benchmarking tool for all backends
+
+This package provides a benchmarking tool for OpenAI-compatible LLM endpoints,
+generating statistics similar to `llama-bench`.
+"""
+
+from ._version import __version__

{llama_benchy-0.1.0 → llama_benchy-0.1.1}/src/llama_benchy/__main__.py

@@ -19,12 +19,13 @@ from transformers import AutoTokenizer
 import requests
 
 # Build number is now imported from __init__.py
-from . import __build__, __version__
+from . import __version__
 
 
 
 def parse_arguments():
     parser = argparse.ArgumentParser(description="LLM Benchmark Script")
+    parser.add_argument("--version", action="version", version=f"%(prog)s {__version__}")
     parser.add_argument("--base-url", type=str, required=True, help="OpenAI compatible endpoint URL")
     parser.add_argument("--api-key", type=str, default="EMPTY", help="API Key for the endpoint")
     parser.add_argument("--model", type=str, required=True, help="Model name to use for benchmarking")
@@ -369,11 +370,10 @@ async def main_async():
         print("Error: --enable-prefix-caching and --no-cache are incompatible.")
         return
 
-    build_number = __build__
     version_number = __version__
 
     current_time = datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")
-    print(f"llama-benchy ({version_number}.{build_number})")
+    print(f"llama-benchy ({version_number})")
     print(f"Date: {current_time}")
     print(f"Benchmarking model: {args.model} at {args.base_url}")
     
@@ -522,7 +522,7 @@ async def main_async():
             print("No results collected. Check if the model is generating tokens.")
         else:
             print(tabulate(results, headers=["model", "test", "t/s", "ttfr (ms)", "est_ppt (ms)", "e2e_ttft (ms)"], tablefmt="pipe", colalign=("left", "right", "right", "right", "right", "right")))
-            print(f"\nllama-benchy ({version_number}.{build_number})")
+            print(f"\nllama-benchy ({version_number})")
             print(f"date: {current_time} | latency mode: {args.latency_mode}")
 
 

{llama_benchy-0.1.0 → llama_benchy-0.1.1}/src/llama_benchy/_version.py

@@ -28,7 +28,7 @@ version_tuple: VERSION_TUPLE
 commit_id: COMMIT_ID
 __commit_id__: COMMIT_ID
 
-__version__ = version = '0.1.0'
-__version_tuple__ = version_tuple = (0, 1, 0)
+__version__ = version = '0.1.1'
+__version_tuple__ = version_tuple = (0, 1, 1)
 
 __commit_id__ = commit_id = None

llama_benchy-0.1.0/src/llama_benchy/__init__.py

@@ -1,24 +0,0 @@
-"""
-llama-benchy - llama-bench style benchmarking tool for all backends
-
-This package provides a benchmarking tool for OpenAI-compatible LLM endpoints,
-generating statistics similar to `llama-bench`.
-"""
-
-from ._version import __version__
-
-# Extract build number from the version string
-# Version format is like: '0.1.dev34+g33f03d886.d20260105'
-# We want to extract the git hash part: '33f03d886'
-__build__ = "unknown"
-if "+" in __version__:
-    try:
-        # Extract the part after the '+' and before the '.'
-        build_part = __version__.split("+")[1].split(".")[0]
-        # Remove the 'g' prefix if it exists
-        if build_part.startswith("g"):
-            __build__ = build_part[1:]
-        else:
-            __build__ = build_part
-    except (IndexError, AttributeError):
-        pass