feat: Add script to automate README benchmark table generation and update max context benchmarks with new models and a kernel parameter change.
这个提交包含在:
@@ -36,10 +36,12 @@ View full benchmarks at: [https://kyuz0.github.io/amd-strix-halo-vllm-toolboxes/
|
||||
| **`meta-llama/Meta-Llama-3.1-8B-Instruct`** | 1 | 128k (0.95) | 128k (0.95) | 128k (0.95) | 128k (0.95) |
|
||||
| **`google/gemma-3-12b-it`** | 1 | 128k (0.95) | 128k (0.95) | 128k (0.95) | 128k (0.95) |
|
||||
| **`openai/gpt-oss-20b`** | 1 | 128k (0.95) | 128k (0.95) | 128k (0.95) | 128k (0.95) |
|
||||
| **`Qwen/Qwen3-14B-AWQ`** | 1 | 40k (0.90) | 40k (0.90) | 40k (0.90) | 40k (0.90) |
|
||||
| **`cpatonn/Qwen3-Coder-30B-A3B-Instruct-GPTQ-4bit`** | 1 | 256k (0.95) | 204k (0.90) | - | - |
|
||||
| **`dazipe/Qwen3-Next-80B-A3B-Instruct-GPTQ-Int4A16`** | 1 | 256k (0.90) | - | - | - |
|
||||
| **`Qwen/Qwen3-14B-AWQ`** | 1 | 40k (0.95) | 40k (0.95) | 40k (0.95) | 40k (0.95) |
|
||||
| **`btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-4bit`** | 1 | 256k (0.95) | 256k (0.95) | 256k (0.95) | 256k (0.95) |
|
||||
| **`btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-8bit`** | 1 | 256k (0.95) | 256k (0.95) | 256k (0.95) | 256k (0.95) |
|
||||
| **`dazipe/Qwen3-Next-80B-A3B-Instruct-GPTQ-Int4A16`** | 1 | 256k (0.95) | 256k (0.95) | 256k (0.95) | 256k (0.95) |
|
||||
| **`openai/gpt-oss-120b`** | 1 | 128k (0.95) | 128k (0.95) | 128k (0.95) | 128k (0.95) |
|
||||
| **`zai-org/GLM-4.7-Flash`** | 1 | 198k (0.95) | 198k (0.95) | 198k (0.95) | 198k (0.95) |
|
||||
|
||||
|
||||
---
|
||||
@@ -184,7 +186,7 @@ amd_iommu=pt amdgpu.gttsize=126976 ttm.pages_limit=32505856
|
||||
|
||||
| Parameter | Purpose |
|
||||
|-----------------------------|--------------------------------------------------------------------------------------------|
|
||||
| `amd_iommu=pt` | Sets IOMMU to pass-through mode; reduces DMA overhead for better performance |
|
||||
| `amd_iommu=off` | Disables AMD IOMMU to reduce overhead for better performance |
|
||||
| `amdgpu.gttsize=126976` | Caps GPU unified memory to 124 GiB; 126976 MiB ÷ 1024 = 124 GiB |
|
||||
| `ttm.pages_limit=32505856` | Caps pinned memory to 124 GiB; 32505856 × 4 KiB = 126976 MiB = 124 GiB |
|
||||
|
||||
|
||||
+83
@@ -0,0 +1,83 @@
|
||||
#!/usr/bin/env python3
|
||||
import json
|
||||
import math
|
||||
from pathlib import Path
|
||||
|
||||
# Config
|
||||
RESULTS_FILE = Path(__file__).parent.parent / "benchmarks/max_context_results.json"
|
||||
|
||||
ORDER = [
|
||||
"meta-llama/Meta-Llama-3.1-8B-Instruct",
|
||||
"google/gemma-3-12b-it",
|
||||
"openai/gpt-oss-20b",
|
||||
"Qwen/Qwen3-14B-AWQ",
|
||||
"btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-4bit",
|
||||
"btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-8bit",
|
||||
"dazipe/Qwen3-Next-80B-A3B-Instruct-GPTQ-Int4A16",
|
||||
"openai/gpt-oss-120b",
|
||||
"zai-org/GLM-4.7-Flash"
|
||||
]
|
||||
|
||||
def format_tokens(n):
|
||||
if n >= 1024:
|
||||
return f"{int(n/1024)}k"
|
||||
return str(n)
|
||||
|
||||
def main():
|
||||
if not RESULTS_FILE.exists():
|
||||
print(f"Error: {RESULTS_FILE} not found.")
|
||||
return
|
||||
|
||||
with open(RESULTS_FILE, "r") as f:
|
||||
data = json.load(f)
|
||||
|
||||
# Organize data: model -> tp -> requests -> result
|
||||
models = {}
|
||||
|
||||
for entry in data:
|
||||
if entry["status"] != "success":
|
||||
continue
|
||||
|
||||
model = entry["model"]
|
||||
tp = entry["tp"]
|
||||
seqs = entry["max_seqs"]
|
||||
util = float(entry["util"])
|
||||
ctx = entry["max_context_1_user"]
|
||||
|
||||
if model not in models:
|
||||
models[model] = {}
|
||||
|
||||
if tp not in models[model]:
|
||||
models[model][tp] = {}
|
||||
|
||||
# Store tuple (ctx, util)
|
||||
# If multiple entries for same seqs (e.g. diff utils), pick standard logic?
|
||||
# The JSON usually has the best working one or we filter.
|
||||
# Assuming unique best entry per seqs/tp tuple from the finder script behavior.
|
||||
models[model][tp][seqs] = (ctx, util)
|
||||
|
||||
# Generate Table
|
||||
print("| Model | TP | 1 Req | 4 Reqs | 8 Reqs | 16 Reqs |")
|
||||
print("| :--- | :--- | :--- | :--- | :--- | :--- |")
|
||||
|
||||
for model_name in ORDER:
|
||||
if model_name not in models:
|
||||
# Identify if there's a different naming or just missing
|
||||
continue
|
||||
|
||||
tps = sorted(models[model_name].keys())
|
||||
for tp in tps:
|
||||
row = [f"**`{model_name}`**", str(tp)]
|
||||
|
||||
for req in [1, 4, 8, 16]:
|
||||
val = models[model_name][tp].get(req)
|
||||
if val:
|
||||
ctx, util = val
|
||||
row.append(f"{format_tokens(ctx)} ({util:.2f})")
|
||||
else:
|
||||
row.append("-")
|
||||
|
||||
print("| " + " | ".join(row) + " |")
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
在新工单中引用
屏蔽一个用户