updates
此提交包含在：
Donato Capitella
+4
@@ -154,6 +154,10 @@ RUN chmod -R a+rwX /opt && \
|
||||
COPY scripts/01-rocm-env-for-triton.sh /etc/profile.d/01-rocm-env-for-triton.sh
|
||||
COPY scripts/99-toolbox-banner.sh /etc/profile.d/99-toolbox-banner.sh
|
||||
COPY scripts/zz-venv-last.sh /etc/profile.d/zz-venv-last.sh
|
||||
COPY scripts/start_vllm.py /usr/local/bin/start-vllm
|
||||
COPY benchmarks/max_context_results.json /opt/max_context_results.json
|
||||
COPY benchmarks/run_vllm_bench.py /opt/run_vllm_bench.py
|
||||
RUN chmod 0644 /etc/profile.d/*.sh && chmod +x /usr/local/bin/start-vllm && chmod 0644 /opt/max_context_results.json
|
||||
RUN chmod 0644 /etc/profile.d/*.sh
|
||||
RUN printf 'ulimit -S -c 0\n' > /etc/profile.d/90-nocoredump.sh && chmod 0644 /etc/profile.d/90-nocoredump.sh
|
||||
|
||||
|
||||
@@ -1,53 +1,40 @@
|
||||
# AMD Strix Halo — vLLM Toolbox/Container (gfx1151, PyTorch + AOTriton)
|
||||
# AMD Strix Halo (gfx1151) — vLLM Toolbox/Container
|
||||
|
||||
An **Arch-based** Docker/Podman container that is **Toolbx-compatible** (usable as a Fedora toolbox) for serving LLMs with **vLLM** on **AMD Ryzen AI Max “Strix Halo” (gfx1151)**. Built on the PyTorch + AOTriton base to make ROCm on Strix Halo practical for day‑to‑day use.
|
||||
An **Arch-based** Docker/Podman container that is **Toolbx-compatible** (usable as a Fedora toolbox) for serving LLMs with **vLLM** on **AMD Ryzen AI Max “Strix Halo” (gfx1151)**. Built on the **TheRock nightly builds** for ROCm.
|
||||
|
||||
> **Built on:** [https://github.com/kyuz0/amd-strix-halo-pytorch-gfx1151-aotriton](https://github.com/kyuz0/amd-strix-halo-pytorch-gfx1151-aotriton)
|
||||
> **Credits:** **lhl** (build tools/scripts), **ssweens** (Arch‑based Dockerfiles), and the **AMD Strix Halo Home Lab Discord** for testing/support.
|
||||
|
||||
---
|
||||
|
||||
## ⚠️ Status & Expectations (Experimental)
|
||||
|
||||
This setup is **highly experimental** on ROCm/Strix Halo. Some models work; **many fail** due to missing custom kernels, unsupported quant types, or TorchInductor/AOTriton limitations on gfx1151. The matrix below lists combinations tested so far. **Please contribute fixes** or additional working recipes (see *Contributing*).
|
||||
|
||||
---
|
||||
|
||||
## Tested Models (Experimental Matrix)
|
||||
|
||||
> **Legend:** ✅ Works (with flags) · ❌ Fails · ⚠️ Notes include the *exact* error/symptom seen.
|
||||
|
||||
| Model (Hugging Face) | Params / Quant | Status | Required flags (if any) | Notes / Errors |
|
||||
| ---------------------------------- | -------------- | -------------------: | ---------------------------------------------------- | ---------------------------------------------------------------------------------------------------- |
|
||||
| `Qwen/Qwen2.5-7B-Instruct` | 7B FP16 | ✅ Works | (recommended) `--dtype float16` | Good baseline; simple serve works. |
|
||||
| `meta-llama/Llama-2-7b-chat-hf` | 7B FP16 | ✅ Works | (recommended) `--dtype float16` | Stable. |
|
||||
| `Qwen/Qwen3-30B-A3B-Instruct-2507` | 30B (A3B) FP16 | ✅ Works | (recommended) `--dtype float16` | |
|
||||
| `Google/Gemma3-27B-Instruct` | 27B FP16 | ✅ Works | (recommended) `--dtype float16` | Slow |
|
||||
| `Google/Gemma3-12B-Instruct` | 12B FP16 | ✅ Works | (recommended) `--dtype float16` | |
|
||||
| `Google/Gemma3-4B-Instruct` |4B FP16 | ✅ Works | (recommended) `--dtype float16` | |
|
||||
| `Qwen/Qwen3-14B-AWQ` | 14B AWQ | ✅ Works (with flags) | `--quantization awq --dtype float16 --enforce-eager` | On ROCm, eager avoids missing `awq_dequantize` during compile; vLLM auto‑sets `VLLM_USE_TRITON_AWQ`. |
|
||||
| `openai/gpt-oss-20b` | 20B MXFP4 | ❌ Fails | — | `ModuleNotFoundError: triton_kernels.matmul_ogs` (MXFP4 path not available in this image). |
|
||||
| `zai-org/GLM-4.5-Air-FP8` | FP8 | ❌ Fails | — | `ValueError: type fp8e4nv not supported (only 'fp8e5')`. |
|
||||
| `cpatonn/GLM-4.5-Air-AWQ-4bit` | AWQ-4bit (MoE) | ❌ Fails | — | Missing custom op: `torch.ops._C.gptq_marlin_repack` (Marlin kernels). |
|
||||
|
||||
> If you get a model to work, please PR a new row with: **model name**, **exact flags**, vLLM version, `torch` & `triton` versions, and a note on **gfx1151** driver/kernel stack.
|
||||
|
||||
---
|
||||
|
||||
## Table of Contents
|
||||
|
||||
* [Tested Models (Benchmarks)](#tested-models-benchmarks)
|
||||
* [1) Toolbx vs Docker/Podman](#1-toolbx-vs-dockerpodman)
|
||||
* [2) Quickstart — Fedora Toolbx (development)](#2-quickstart--fedora-toolbx-development)
|
||||
* [3) Testing the API](#3-testing-the-api)
|
||||
* [4) Quickstart — Podman/Docker](#4-quickstart--podmandocker)
|
||||
* [5) Models, dtypes & storage](#5-models-dtypes--storage)
|
||||
* [6) Performance notes (short)](#6-performance-notes-short)
|
||||
* [7) Requirements (host)](#7-requirements-host)
|
||||
* [8) Acknowledgements & Links](#8-acknowledgements--links)
|
||||
* [Tested Models](#tested-models)
|
||||
* [Contributing](#contributing)
|
||||
* [3) Quickstart — Ubuntu (Distrobox)](#3-quickstart--ubuntu-distrobox)
|
||||
* [4) Testing the API](#4-testing-the-api)
|
||||
* [5) Use a Web UI for Chatting](#5-use-a-web-ui-for-chatting)
|
||||
|
||||
|
||||
## Tested Models (Benchmarks)
|
||||
|
||||
View full benchmarks at: [https://kyuz0.github.io/amd-strix-halo-vllm-toolboxes/](https://kyuz0.github.io/amd-strix-halo-vllm-toolboxes/)
|
||||
|
||||
**Table Key:** Cell values represent `Max Context Length (GPU Memory Utilization)`.
|
||||
|
||||
| Model | TP | 1 Req | 4 Reqs | 8 Reqs | 16 Reqs |
|
||||
| :--- | :--- | :--- | :--- | :--- | :--- |
|
||||
| **`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) | - | - | - |
|
||||
| **`openai/gpt-oss-120b`** | 1 | 128k (0.95) | 128k (0.95) | 128k (0.95) | 128k (0.95) |
|
||||
|
||||
|
||||
---
|
||||
|
||||
## 1) Toolbx vs Docker/Podman
|
||||
|
||||
The `kyuz0/vllm-therock-gfx1151-aotriton:latest` image can be used both as:
|
||||
@@ -63,7 +50,7 @@ Create a toolbox that exposes the GPU and relaxes seccomp to avoid ROCm syscall
|
||||
|
||||
```bash
|
||||
toolbox create vllm \
|
||||
--image docker.io/kyuz0/vllm-therock-gfx1151-aotriton:latest \
|
||||
--image docker.io/kyuz0/vllm-therock-gfx1151:latest \
|
||||
-- --device /dev/dri --device /dev/kfd \
|
||||
--group-add video --group-add render --security-opt seccomp=unconfined
|
||||
```
|
||||
@@ -74,33 +61,45 @@ Enter it:
|
||||
toolbox enter vllm
|
||||
```
|
||||
|
||||
**Model storage (Toolbx):** keep weights **outside** the toolbox under your HOME so they persist. Recommended path:
|
||||
**Model storage:** Models are downloaded to `~/.cache/huggingface` by default. This directory is shared with the host if you created the toolbox correctly, so downloads persist.
|
||||
|
||||
```bash
|
||||
mkdir -p ~/vllm-models
|
||||
```
|
||||
### Serving a Model (Easiest Way)
|
||||
|
||||
Serve a model using the helper script **`start-vllm`** (it prints the exact `vllm serve` command and then runs it). Models download to `~/vllm-models` by default; if a model isn't present, it will be fetched from Hugging Face automatically:
|
||||
The toolbox includes a TUI wizard called **`start-vllm`** which includes pre-configured models and handles the launch flags for you. This is the easiest way to get started.
|
||||
|
||||
```bash
|
||||
start-vllm
|
||||
# pick a model from the menu; the script prints the serve command and launches it
|
||||
```
|
||||
|
||||
> Defaults: `0.0.0.0:8000` and `~/vllm-models` for weights. You can still run `vllm serve` manually if you prefer.
|
||||
|
||||
> Toolbx shares HOME by design, so `~/vllm-models` stays on the host and survives toolbox updates.
|
||||
>
|
||||
> **Cache note (Toolbx):** vLLM will also write compiled kernels to `~/.cache/vllm/torch_compile_cache/` in your HOME. For example:
|
||||
>
|
||||
> ```bash
|
||||
> du -sh ~/.cache/vllm/torch_compile_cache/
|
||||
> # e.g., 138M /home/you/.cache/vllm/torch_compile_cache/
|
||||
> ```
|
||||
> **Cache note:** vLLM writes compiled kernels to `~/.cache/vllm/`.
|
||||
|
||||
---
|
||||
|
||||
## 3) Testing the API
|
||||
## 3) Quickstart — Ubuntu (Distrobox)
|
||||
|
||||
Ubuntu’s toolbox package still breaks GPU access, so use Distrobox instead:
|
||||
|
||||
```bash
|
||||
distrobox create -n vllm \
|
||||
--image docker.io/kyuz0/vllm-therock-gfx1151:latest \
|
||||
--additional-flags "--device /dev/kfd --device /dev/dri --group-add video --group-add render --security-opt seccomp=unconfined"
|
||||
|
||||
distrobox enter vllm
|
||||
```
|
||||
|
||||
> **Verification:** Run `rocm-smi` to check GPU status.
|
||||
|
||||
### Serving a Model (Easiest Way)
|
||||
|
||||
The toolbox includes a TUI wizard called **`start-vllm`** which includes pre-configured models and handles the launch flags for you. This is the easiest way to get started.
|
||||
|
||||
```bash
|
||||
start-vllm
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 4) Testing the API
|
||||
|
||||
Once the server is up, hit the OpenAI‑compatible endpoint:
|
||||
|
||||
@@ -125,145 +124,21 @@ MODEL=$(curl -s http://localhost:8000/v1/models | jq -r '.data[0].id') curl -X P
|
||||
|
||||
---
|
||||
|
||||
## 4) Quickstart — Podman/Docker
|
||||
## 5) Use a Web UI for Chatting
|
||||
|
||||
Prefer this for persistent services. **Always mount a host directory for weights** so they live outside the container. If the model isn't present, vLLM will fetch it from **Hugging Face** into the mapped directory.
|
||||
|
||||
**Qwen2.5 7B Instruct**
|
||||
If vLLM is on a remote server, expose port 8000 via SSH port forwarding:
|
||||
|
||||
```bash
|
||||
podman run -d --name vllm-qwen2p5-7b \
|
||||
--ipc=host \
|
||||
--network host \
|
||||
--device /dev/kfd \
|
||||
--device /dev/dri \
|
||||
--group-add video \
|
||||
--group-add render \
|
||||
-v ~/vllm-models:/models \
|
||||
-v ~/.cache/vllm:/root/.cache/vllm \
|
||||
docker.io/kyuz0/vllm-therock-gfx1151-aotriton:latest \
|
||||
bash -lc 'source /torch-therock/.venv/bin/activate; \
|
||||
TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 \
|
||||
vllm serve Qwen/Qwen2.5-7B-Instruct --dtype float16 \
|
||||
--host 0.0.0.0 --port 8000 --download-dir /models'
|
||||
ssh -L 0.0.0.0:8000:localhost:8000 <vllm-host>
|
||||
```
|
||||
|
||||
> Not using `--network host`? Map a port instead: `-p 8000:8000`.
|
||||
|
||||
For other models, you can try:
|
||||
|
||||
|
||||
**Qwen3 30B A3B Instruct (2507)**
|
||||
Then, you can start HuggingFace ChatUI like this (on your host):
|
||||
|
||||
```bash
|
||||
podman run -d --name vllm-qwen3-30b-a3b \
|
||||
--ipc=host \
|
||||
--network host \
|
||||
--device /dev/kfd \
|
||||
--device /dev/dri \
|
||||
--group-add video \
|
||||
--group-add render \
|
||||
-v ~/vllm-models:/models \
|
||||
-v ~/.cache/vllm:/root/.cache/vllm \
|
||||
docker.io/kyuz0/vllm-therock-gfx1151-aotriton:latest \
|
||||
bash -lc 'source /torch-therock/.venv/bin/activate; \
|
||||
TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 \
|
||||
vllm serve Qwen/Qwen3-30B-A3B-Instruct-2507 --dtype float16 \
|
||||
--host 0.0.0.0 --port 8000 --download-dir /models'
|
||||
docker run -p 3000:3000 \
|
||||
--add-host=host.docker.internal:host-gateway \
|
||||
-e OPENAI_BASE_URL=http://host.docker.internal:8000/v1 \
|
||||
-e OPENAI_API_KEY=dummy \
|
||||
-v chat-ui-data:/data \
|
||||
ghcr.io/huggingface/chat-ui-db
|
||||
```
|
||||
|
||||
**Qwen3 14B AWQ** *(requires extra flags on ROCm)*
|
||||
|
||||
```bash
|
||||
podman run -d --name vllm-qwen3-14b-awq \
|
||||
--ipc=host \
|
||||
--network host \
|
||||
--device /dev/kfd \
|
||||
--device /dev/dri \
|
||||
--group-add video \
|
||||
--group-add render \
|
||||
-v ~/vllm-models:/models \
|
||||
-v ~/.cache/vllm:/root/.cache/vllm \
|
||||
docker.io/kyuz0/vllm-therock-gfx1151-aotriton:latest \
|
||||
bash -lc 'source /torch-therock/.venv/bin/activate; \
|
||||
TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 \
|
||||
vllm serve Qwen/Qwen3-14B-AWQ --quantization awq --dtype float16 --enforce-eager \
|
||||
--host 0.0.0.0 --port 8000 --download-dir /models'
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 5) Models, dtypes & storage
|
||||
|
||||
* Start with **Qwen/Qwen2.5-7B-Instruct**; larger models may work but are less forgiving on unified memory.
|
||||
* Use `--dtype float16` unless you have a reason to change.
|
||||
* **Storage discipline:**
|
||||
|
||||
* **Toolbx:** `--download-dir ~/vllm-models` (lives in your HOME on the host).
|
||||
* **Podman/Docker:** `-v ~/vllm-models:/models` and `--download-dir /models`.
|
||||
|
||||
---
|
||||
|
||||
## 6) Performance notes (short)
|
||||
|
||||
* The image is built on the PyTorch + **AOTriton** base; enabling `TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1` can improve startup/throughput on some models.
|
||||
* vLLM flags you might tune later: `--gpu-memory-utilization`, `--max-num-seqs`, `--max-model-len`. Start simple; add knobs only if needed.
|
||||
|
||||
---
|
||||
|
||||
## 7) Requirements (host)
|
||||
|
||||
**Hardware & drivers**
|
||||
|
||||
* AMD Strix Halo APU (gfx1151).
|
||||
* Working amdgpu stack with `/dev/kfd` (ROCm compute) and `/dev/dri` (graphics).
|
||||
* Your user in the **video** and **render** groups.
|
||||
|
||||
**Unified memory setup (HIGHLY recommended)**
|
||||
Enable large GTT/unified memory so the iGPU can borrow system RAM for bigger models:
|
||||
|
||||
1. **Kernel parameters** (append to your GRUB cmdline):
|
||||
|
||||
```
|
||||
amd_iommu=off amdgpu.gttsize=131072 ttm.pages_limit=33554432
|
||||
```
|
||||
|
||||
| Parameter | Purpose |
|
||||
| -------------------------- | ---------------------------- |
|
||||
| `amd_iommu=off` | Reduces latency |
|
||||
| `amdgpu.gttsize=131072` | 128 GiB GTT (unified memory) |
|
||||
| `ttm.pages_limit=33554432` | Large pinned allocations |
|
||||
|
||||
2. **BIOS**: allocate **minimal VRAM** to the iGPU (e.g., **512 MB**) and rely on unified memory.
|
||||
|
||||
3. **Fedora example** (GRUB): edit `/etc/default/grub` → `GRUB_CMDLINE_LINUX=...` then:
|
||||
|
||||
```bash
|
||||
sudo grub2-mkconfig -o /boot/grub2/grub.cfg
|
||||
sudo reboot
|
||||
```
|
||||
|
||||
**Container runtime**
|
||||
|
||||
* Podman or Docker installed (examples use Podman; replace with Docker if preferred).
|
||||
|
||||
---
|
||||
|
||||
## 8) Contributing
|
||||
|
||||
Spotted a fix, a working flag combo, or a model that should be on the list? **PRs welcome!** Please include:
|
||||
|
||||
* Model repo + exact version tag (if any)
|
||||
* Full `vllm serve` command/flags that work
|
||||
* vLLM version, `torch` & `triton` versions (`python -c "import torch, triton; print(torch.__version__, triton.__version__)"`)
|
||||
* Short log snippet of success/failure (especially the **first** error)
|
||||
* Any relevant kernel/AOTriton env vars (e.g., `TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1`)
|
||||
|
||||
---
|
||||
|
||||
## 9) Acknowledgements & Links
|
||||
|
||||
* Base images & docs: [https://github.com/kyuz0/amd-strix-halo-pytorch-gfx1151-aotriton](https://github.com/kyuz0/amd-strix-halo-pytorch-gfx1151-aotriton)
|
||||
* Upstreams: [vLLM](https://github.com/vllm-project/vllm), [ROCm/TheRock](https://github.com/ROCm/TheRock), [AOTriton](https://github.com/ROCm/aotriton)
|
||||
* Community: **AMD Strix Halo Home Lab Discord** — [https://discord.gg/pnPRyucNrG](https://discord.gg/pnPRyucNrG)
|
||||
* Big thanks to **lhl** and **ssweens** for doing the actual heavy lifting for this.
|
||||
|
||||
@@ -0,0 +1,7 @@
|
||||
{
|
||||
"elapsed_time": 1302.7062463890015,
|
||||
"num_requests": 200,
|
||||
"total_num_tokens": 146805,
|
||||
"requests_per_second": 0.15352655332265747,
|
||||
"tokens_per_second": 112.69232830266365
|
||||
}
|
||||
+7
@@ -0,0 +1,7 @@
|
||||
{
|
||||
"elapsed_time": 540.2676798280002,
|
||||
"num_requests": 200,
|
||||
"total_num_tokens": 146805,
|
||||
"requests_per_second": 0.37018686748700586,
|
||||
"tokens_per_second": 271.7264154071495
|
||||
}
|
||||
+7
@@ -0,0 +1,7 @@
|
||||
{
|
||||
"elapsed_time": 1303.4944151099999,
|
||||
"num_requests": 200,
|
||||
"total_num_tokens": 146805,
|
||||
"requests_per_second": 0.15343372221746138,
|
||||
"tokens_per_second": 112.62418795067208
|
||||
}
|
||||
@@ -0,0 +1,7 @@
|
||||
{
|
||||
"elapsed_time": 914.8563823220001,
|
||||
"num_requests": 200,
|
||||
"total_num_tokens": 148857,
|
||||
"requests_per_second": 0.21861354838273012,
|
||||
"tokens_per_second": 162.71078485804028
|
||||
}
|
||||
@@ -0,0 +1,7 @@
|
||||
{
|
||||
"elapsed_time": 522.8661062630126,
|
||||
"num_requests": 200,
|
||||
"total_num_tokens": 145877,
|
||||
"requests_per_second": 0.38250710383471637,
|
||||
"tokens_per_second": 278.99494393048457
|
||||
}
|
||||
@@ -0,0 +1,7 @@
|
||||
{
|
||||
"elapsed_time": 1339.915984058,
|
||||
"num_requests": 200,
|
||||
"total_num_tokens": 147036,
|
||||
"requests_per_second": 0.14926308990977954,
|
||||
"tokens_per_second": 109.73523843987172
|
||||
}
|
||||
@@ -0,0 +1,7 @@
|
||||
{
|
||||
"elapsed_time": 468.4791132300161,
|
||||
"num_requests": 200,
|
||||
"total_num_tokens": 147036,
|
||||
"requests_per_second": 0.42691337639593563,
|
||||
"tokens_per_second": 313.85817605876395
|
||||
}
|
||||
@@ -0,0 +1,575 @@
|
||||
#!/usr/bin/env python3
|
||||
import subprocess
|
||||
import time
|
||||
import socket
|
||||
import json
|
||||
import sys
|
||||
import os
|
||||
import requests
|
||||
import re
|
||||
import argparse
|
||||
from pathlib import Path
|
||||
try:
|
||||
from transformers import AutoConfig
|
||||
except ImportError:
|
||||
print("Error: 'transformers' not found. Please install it or run in vLLM environment.")
|
||||
sys.exit(1)
|
||||
|
||||
# Import configuration from average benchmark script
|
||||
try:
|
||||
from run_vllm_bench import MODEL_TABLE, MODELS_TO_RUN, get_gpu_count, kill_vllm
|
||||
except ImportError:
|
||||
print("Error: Could not import run_vllm_bench.py. Make sure it is in the same directory.")
|
||||
sys.exit(1)
|
||||
|
||||
# =========================
|
||||
# 🧠 GROUNDING & METHODOLOGY
|
||||
# =========================
|
||||
# This script finds the Maximum Working Context (MWC) for vLLM models.
|
||||
#
|
||||
# Methodology:
|
||||
# 1. **Inspect**: Use `transformers.AutoConfig` to determine the model's theoretical limit
|
||||
# (e.g., `max_position_embeddings`).
|
||||
# 2. **Probe**: Launch `vllm serve` at this limit.
|
||||
# 3. **React**:
|
||||
# - If stable ("Application startup complete"): Success.
|
||||
# - If OOM ("KV cache capacity... is X"): Retry with vLLM's suggested X.
|
||||
# - If Config Error ("max_model_len... is Y"): Retry with vLLM's suggested Y.
|
||||
|
||||
# =========================
|
||||
# ⚙️ CONFIG
|
||||
# =========================
|
||||
HOST = "127.0.0.1"
|
||||
PORT = 8000
|
||||
RESULTS_FILE = Path("max_context_results.json")
|
||||
REPORT_FILE = Path("max_context_report.md")
|
||||
|
||||
# We test these GPU Utilizations steps to see how much we can squeeze
|
||||
# 0.90 is default, but we want MAX context.
|
||||
# 0.98 is our target high. 0.95 is the fallback.
|
||||
GPU_UTIL_STEPS = ["0.95", "0.90"]
|
||||
# We test these concurrency settings
|
||||
CONCURRENCY_STEPS = [1, 4, 8, 16]
|
||||
|
||||
def log(msg): print(f"[MAX-CTX] {msg}", flush=True)
|
||||
|
||||
def get_hf_context_limit(model_name, trust_remote=False):
|
||||
try:
|
||||
cfg = AutoConfig.from_pretrained(model_name, trust_remote_code=trust_remote)
|
||||
|
||||
# Gemma 3 and similar multi-config models
|
||||
if hasattr(cfg, "text_config"):
|
||||
tc = cfg.text_config
|
||||
if hasattr(tc, "max_position_embeddings"):
|
||||
return int(tc.max_position_embeddings)
|
||||
|
||||
# Standard HF attributes
|
||||
for attr in (
|
||||
"max_position_embeddings",
|
||||
"seq_length",
|
||||
"max_seq_len",
|
||||
"n_positions",
|
||||
):
|
||||
val = getattr(cfg, attr, None)
|
||||
if val is not None:
|
||||
return int(val)
|
||||
|
||||
return 8192
|
||||
|
||||
except Exception as e:
|
||||
log(f"Warning: Could not read config for {model_name}: {e}. Defaulting to 32768.")
|
||||
return 32768
|
||||
|
||||
def get_vllm_server_cmd(model, tp_size, util, max_len, max_seqs):
|
||||
"""
|
||||
Constructs the vLLM serve command.
|
||||
"""
|
||||
config = MODEL_TABLE[model]
|
||||
|
||||
cmd = [
|
||||
"vllm", "serve", model,
|
||||
"--gpu-memory-utilization", str(util),
|
||||
"--max-model-len", str(max_len),
|
||||
"--tensor-parallel-size", str(tp_size),
|
||||
"--max-num-seqs", str(max_seqs),
|
||||
"--dtype", "auto",
|
||||
# "--disable-log-stats" # Cleaner output, but user managed without it
|
||||
]
|
||||
|
||||
if config.get("trust_remote"): cmd.append("--trust-remote-code")
|
||||
if config.get("enforce_eager"): cmd.append("--enforce-eager")
|
||||
|
||||
# Add model specific env vars
|
||||
env = os.environ.copy()
|
||||
env.update(config.get("env", {}))
|
||||
|
||||
return cmd, env
|
||||
|
||||
def is_port_free(port):
|
||||
with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
|
||||
return s.connect_ex(('localhost', port)) != 0
|
||||
|
||||
def force_cleanup(hard=False):
|
||||
"""
|
||||
Kills vLLM using multiple methods and ensures port is free.
|
||||
BLOCKS until processes are definitely gone.
|
||||
"""
|
||||
timeout = 20 if hard else 10
|
||||
start_time = time.time()
|
||||
|
||||
while True:
|
||||
# 1. Aggressive Kill Commands
|
||||
# We send these EVERY loop iteration until they die.
|
||||
subprocess.run("pkill -9 -f 'vllm.entrypoints.api_server'", shell=True, stderr=subprocess.DEVNULL)
|
||||
subprocess.run("pkill -9 -f 'vllm serve'", shell=True, stderr=subprocess.DEVNULL)
|
||||
subprocess.run("pkill -9 -f 'VLLM::'", shell=True, stderr=subprocess.DEVNULL)
|
||||
subprocess.run("pkill -9 -f 'multiprocessing.spawn'", shell=True, stderr=subprocess.DEVNULL)
|
||||
subprocess.run("pkill -9 -f ray::", shell=True, stderr=subprocess.DEVNULL)
|
||||
|
||||
# 2. Check if they are still there
|
||||
# We check specifically for the persistence of any vllm-related process
|
||||
# We use explicit list to know WHICH one triggered it
|
||||
# CRITICAL FIX: We MUST use shell=False otherwise 'pgrep -f pattern'
|
||||
# matches the 'sh -c pgrep ... pattern' command content itself!
|
||||
dirty = False
|
||||
|
||||
# Check 1: vllm serve
|
||||
if subprocess.run(["pgrep", "-f", "vllm serve"], stdout=subprocess.DEVNULL).returncode == 0:
|
||||
# Double check it's not us (Python script)
|
||||
# But simpler to just proceed if we trust shell=False works
|
||||
log("Clean waiting: Found 'vllm serve' process:")
|
||||
subprocess.run("pgrep -a -f 'vllm serve'", shell=True) # debug
|
||||
dirty = True
|
||||
|
||||
# Check 2: api_server
|
||||
if subprocess.run(["pgrep", "-f", "vllm.entrypoints.api_server"], stdout=subprocess.DEVNULL).returncode == 0:
|
||||
log("Clean waiting: Found 'vllm.entrypoints.api_server' process:")
|
||||
subprocess.run("pgrep -a -f 'vllm.entrypoints.api_server'", shell=True) # debug
|
||||
dirty = True
|
||||
|
||||
# Check 3: VLLM:: (Ray workers)
|
||||
if subprocess.run(["pgrep", "-f", "VLLM::"], stdout=subprocess.DEVNULL).returncode == 0:
|
||||
log("Clean waiting: Found 'VLLM::' process:")
|
||||
subprocess.run("pgrep -a -f 'VLLM::'", shell=True) # debug
|
||||
dirty = True
|
||||
|
||||
if not dirty:
|
||||
# Processes are gone. Now check port.
|
||||
if is_port_free(PORT):
|
||||
time.sleep(1) # Final safety buffer
|
||||
return # Clean!
|
||||
else:
|
||||
log("Clean: Processes gone, but Port 8000 still held. Waiting...")
|
||||
else:
|
||||
log("Clean: Processes still detected. Retrying kill...")
|
||||
|
||||
if time.time() - start_time > timeout:
|
||||
log("CRITICAL: Cleanup timed out! Force attempting `killall -9 vllm` as last resort.")
|
||||
subprocess.run("killall -9 vllm", shell=True, stderr=subprocess.DEVNULL)
|
||||
break
|
||||
|
||||
time.sleep(1.5) # Wait a bit before hammering again
|
||||
|
||||
|
||||
def wait_for_server_and_parse(process, timeout=300):
|
||||
"""
|
||||
Waits for server to be ready.
|
||||
Parses stdout for "Count of GPU blocks" and "Block size".
|
||||
Returns: (ready_bool, gpu_blocks, block_size, max_len_clamped, failure_reason)
|
||||
"""
|
||||
start = time.time()
|
||||
gpu_blocks = 0
|
||||
block_size = 16 # default
|
||||
max_len_clamped = None
|
||||
|
||||
logs = []
|
||||
failure_reason = None
|
||||
|
||||
while time.time() - start < timeout:
|
||||
if process.poll() is not None:
|
||||
# Process died.
|
||||
for line in process.stdout:
|
||||
line_str = line.decode("utf-8", errors="replace").strip()
|
||||
logs.append(line_str)
|
||||
|
||||
# SCAN FULL HISTORY if not found yet
|
||||
# Sometimes error was in previous lines or split
|
||||
if not failure_reason:
|
||||
full_log = "\n".join(logs)
|
||||
|
||||
# Check 1: Sampler OOM
|
||||
if "warming up sampler" in full_log and "CUDA out of memory" in full_log:
|
||||
failure_reason = "Sampler Warmup OOM"
|
||||
|
||||
# Check 2: Explicit vLLM suggestion (Estimated)
|
||||
# "estimated maximum model length is 127120"
|
||||
elif "estimated maximum model length is" in full_log:
|
||||
m = re.search(r"estimated maximum model length is (\d+)", full_log)
|
||||
if m:
|
||||
failure_reason = f"estimated maximum model length is {m.group(1)}"
|
||||
|
||||
# Check 3: Derived Max Model Len
|
||||
# "derived max_model_len (max_position_embeddings=131072.0 ...)"
|
||||
elif "derived max_model_len" in full_log:
|
||||
failure_reason = "derived max_model_len detected"
|
||||
|
||||
# Check 4: Capacity/Value Error
|
||||
elif "ValueError" in full_log and "maximum number of tokens" in full_log:
|
||||
failure_reason = "Capacity Error (Found in history)"
|
||||
|
||||
# Check 5: Generic OOM
|
||||
elif "CUDA out of memory" in full_log or "hipErrorOutOfMemory" in full_log:
|
||||
failure_reason = "OOM detected"
|
||||
|
||||
if not failure_reason:
|
||||
# Unexpected death! Dump logs to see why.
|
||||
log("CRITICAL: Process died unexpectedly! Dumping last 100 lines:")
|
||||
print("=== vLLM SERVER LOGS (LAST 100 LINES) ===")
|
||||
for l in logs[-100:]:
|
||||
print(l)
|
||||
print("=============================================")
|
||||
|
||||
return False, 0, 0, None, failure_reason
|
||||
|
||||
line = process.stdout.readline()
|
||||
if line:
|
||||
line_str = line.decode("utf-8", errors="replace").strip()
|
||||
logs.append(line_str)
|
||||
|
||||
# 1. Parse Legacy "GPU blocks" (if present)
|
||||
m_blocks = re.search(r"# GPU blocks:\s*(\d+)", line_str)
|
||||
if m_blocks:
|
||||
gpu_blocks = int(m_blocks.group(1))
|
||||
block_size = 16 # assume default unless found
|
||||
log(f" -> Found GPU blocks: {gpu_blocks} (Legacy)")
|
||||
|
||||
# 2. Parse Newer "GPU KV cache size" (vLLM 0.11+)
|
||||
# "GPU KV cache size: 111,536 tokens"
|
||||
m_kv_tokens = re.search(r"GPU KV cache size:\s*([\d,]+)\s*tokens", line_str)
|
||||
if m_kv_tokens:
|
||||
tokens_str = m_kv_tokens.group(1).replace(",", "")
|
||||
gpu_blocks = int(tokens_str) # We use 'gpu_blocks' variable to store total tokens now for simplicity
|
||||
block_size = 1 # Effectively 1 because we have the total count
|
||||
log(f" -> Found GPU KV Cache tokens: {gpu_blocks}")
|
||||
|
||||
# 3. Parse Block Size (optional, mostly for legacy)
|
||||
m_bs = re.search(r"block_size=(\d+)", line_str)
|
||||
if m_bs:
|
||||
block_size = int(m_bs.group(1))
|
||||
|
||||
# Failure hints
|
||||
if "ValueError" in line_str and "maximum number of tokens" in line_str:
|
||||
failure_reason = line_str
|
||||
if "derived max_model_len" in line_str:
|
||||
failure_reason = line_str
|
||||
if "warming up sampler" in line_str and "CUDA out of memory" in line_str:
|
||||
failure_reason = "Sampler Warmup OOM"
|
||||
elif "CUDA out of memory" in line_str or "hipErrorOutOfMemory" in line_str:
|
||||
failure_reason = "OOM detected"
|
||||
|
||||
# Check for startup
|
||||
if "Application startup complete" in line_str or "Uvicorn running on" in line_str:
|
||||
if gpu_blocks > 0:
|
||||
log(" -> Server signal detected. Waiting 5s for socket stability...")
|
||||
time.sleep(5)
|
||||
return True, gpu_blocks, block_size, max_len_clamped, None
|
||||
else:
|
||||
return False, 0, 0, None, "Parsed Success but Token/Block Count was 0"
|
||||
|
||||
# Timeout case
|
||||
log("CRITICAL: Server startup timed out! Dumping last 100 lines:")
|
||||
print("=== vLLM SERVER LOGS (LAST 100 LINES) ===")
|
||||
for l in logs[-100:]:
|
||||
print(l)
|
||||
print("=============================================")
|
||||
return False, 0, 0, None, "Timeout"
|
||||
|
||||
def verify_context(model, context_len):
|
||||
"""
|
||||
Sends a request to the server with length ~context_len to verify stability.
|
||||
"""
|
||||
url = f"http://{HOST}:{PORT}/v1/completions"
|
||||
|
||||
# We use a simple "A " * N prompt.
|
||||
# Llama 3 tokenizer: "A" is usually 1 token.
|
||||
|
||||
prompt = "A " * int(context_len * 0.5) # 50% fill to be safe/approx
|
||||
|
||||
payload = {
|
||||
"model": model,
|
||||
"prompt": prompt,
|
||||
"max_tokens": 10,
|
||||
"temperature": 0
|
||||
}
|
||||
|
||||
# Retry loop for connection refusals (race condition)
|
||||
max_retries = 5
|
||||
for attempt in range(max_retries):
|
||||
try:
|
||||
# Increased timeout to 300s because prefilling 60k+ tokens takes time!
|
||||
r = requests.post(url, json=payload, timeout=300)
|
||||
if r.status_code == 200:
|
||||
return True, "Success"
|
||||
else:
|
||||
# If 500 or 400 error, maybe we shouldn't retry? Usually yes for 500 if transient.
|
||||
# But for now let's just fail or retry.
|
||||
# If we are OOMing, we will likely get a 500 or it will hang.
|
||||
return False, f"HTTP {r.status_code}: {r.text[:200]}"
|
||||
except requests.exceptions.ConnectionError:
|
||||
if attempt < max_retries - 1:
|
||||
log(f" -> Connection refused. Retrying verification ({attempt+1}/{max_retries})...")
|
||||
time.sleep(2)
|
||||
else:
|
||||
return False, "Connection Refused (Max Retries)"
|
||||
except Exception as e:
|
||||
return False, str(e)
|
||||
|
||||
return False, "Unknown Error"
|
||||
|
||||
def run_probe(model, tp, util, max_seqs, start_limit=None):
|
||||
"""
|
||||
Probes a specific configuration starting from the model's architectural limit.
|
||||
"""
|
||||
trust_remote = MODEL_TABLE[model].get("trust_remote", False)
|
||||
# 1. Get the Advertised Limit (The "Smart" Way)
|
||||
arch_limit = get_hf_context_limit(model, trust_remote)
|
||||
|
||||
# Intelligent Start: If we know a lower limit worked for lower concurrency, start there.
|
||||
target_len = arch_limit
|
||||
if start_limit:
|
||||
target_len = min(arch_limit, start_limit)
|
||||
log(f" -> Smart Start: Capping initial probe at {target_len} (based on previous run)")
|
||||
|
||||
result_data = {
|
||||
"model": model,
|
||||
"tp": tp,
|
||||
"util": util,
|
||||
"max_seqs": max_seqs,
|
||||
"model_limit": arch_limit,
|
||||
"configured_len": 0,
|
||||
"real_capacity": 0,
|
||||
"status": "fail",
|
||||
"error": ""
|
||||
}
|
||||
|
||||
log(f"Probing {model} | TP={tp} | Util={util} | Seqs={max_seqs} | Model Limit={arch_limit}")
|
||||
|
||||
# We loop until we succeed OR we drop below a useful context size.
|
||||
while target_len >= 2048:
|
||||
force_cleanup()
|
||||
|
||||
cmd, env = get_vllm_server_cmd(model, tp, util, target_len, max_seqs)
|
||||
log(f"DEBUG: Cmd: {' '.join(cmd)}")
|
||||
|
||||
proc = None
|
||||
try:
|
||||
proc = subprocess.Popen(cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, env=env)
|
||||
ready, blocks, block_size, _, fail_msg = wait_for_server_and_parse(proc)
|
||||
|
||||
if ready:
|
||||
# Success - but let's VERIFY it actually answers
|
||||
total_capacity = blocks * block_size
|
||||
workable_len = min(target_len, total_capacity)
|
||||
|
||||
# Verify with actual request
|
||||
# We cap verification at 4096 because we just want to know if it crashes,
|
||||
# we don't need to wait for a 128k context fill just for a liveness check.
|
||||
verify_len = min(workable_len, 4096)
|
||||
log(f" -> Server ready. Verifying stability with approx {int(verify_len * 0.5)} tokens (capped at 4k)...")
|
||||
v_ok, v_msg = verify_context(model, verify_len)
|
||||
|
||||
if v_ok:
|
||||
log(f" -> Success! capacity={total_capacity}, configured={workable_len}")
|
||||
log(f" -> Verification passed: {v_msg}")
|
||||
|
||||
# Cleanup SUCCESSFUL process immediately
|
||||
proc.terminate()
|
||||
try: proc.wait(timeout=5)
|
||||
except: proc.kill()
|
||||
|
||||
result_data["status"] = "success"
|
||||
result_data["configured_len"] = target_len
|
||||
result_data["real_capacity"] = total_capacity
|
||||
result_data["max_context_1_user"] = workable_len
|
||||
|
||||
return result_data
|
||||
else:
|
||||
log(f" -> Server started, but Verification FAILED: {v_msg}")
|
||||
# Treat as a crash/failure, back off
|
||||
fail_msg = "Verification Failed"
|
||||
|
||||
# Capture any remaining logs if the process is dead or dying
|
||||
# Or just read what's currently available non-blocking?
|
||||
# Simpler: just terminate and read output.
|
||||
proc.terminate()
|
||||
try:
|
||||
outs, errs = proc.communicate(timeout=5)
|
||||
if outs:
|
||||
print("=== vLLM SERVER LOGS (DURING VERIFICATION FAILURE) ===")
|
||||
print(outs.decode('utf-8', errors='replace'))
|
||||
print("======================================================")
|
||||
except:
|
||||
proc.kill()
|
||||
|
||||
|
||||
# If we fall through here, ready=False OR verify=False
|
||||
log(f" -> Attempt failed at {target_len}")
|
||||
if fail_msg: log(f" Reason: {fail_msg}")
|
||||
result_data["error"] = fail_msg if fail_msg else "Process died or timed out"
|
||||
|
||||
if fail_msg:
|
||||
# Case V: Verification Failed (Server up, but unstable inference)
|
||||
# User requests drop to 0.95 tier immediately.
|
||||
# Must check this FIRST to ensure we don't fall through.
|
||||
if "Verification Failed" in str(fail_msg):
|
||||
log(" -> Verification Failed (Unstable). Aborting this Util, dropping to lower tier.")
|
||||
break
|
||||
|
||||
# Case S: Sampler Warmup OOM (Fatal for this Util)
|
||||
if "Sampler Warmup OOM" in fail_msg:
|
||||
log(" -> Critical Sampler OOM. Utilization/Seqs too high. Aborting this configuration.")
|
||||
break # Give up on this Util/Seq combo immediately
|
||||
|
||||
# Case X: Dirty State / Zombie VRAM
|
||||
# "Free memory on device (1.56/31.86 GiB) on startup is less than desired..."
|
||||
if "Free memory on device" in fail_msg and "less than desired" in fail_msg:
|
||||
log(" -> Dirty VRAM detected (previous run didn't cleanup?). Retrying with HARD cleanup.")
|
||||
force_cleanup(hard=True)
|
||||
continue # Retry SAME target_len
|
||||
|
||||
# Case A: VRAM Limit ("maximum number of tokens... is X")
|
||||
m_capacity = re.search(r"maximum number of tokens.*?KV cache is (\d+)", fail_msg)
|
||||
if m_capacity:
|
||||
cap = int(m_capacity.group(1))
|
||||
log(f" -> Found Hardware Capacity: {cap}")
|
||||
target_len = cap
|
||||
continue # Retry Exact Cap
|
||||
|
||||
# Case B: Model Limit mismatch
|
||||
# "Value error, User-specified max_model_len (500000) is greater than the derived max_model_len (max_position_embeddings=131072.0 ...)"
|
||||
# We regex for 'derived max_model_len' and then look for numbers in the proximity.
|
||||
|
||||
if "derived max_model_len" in fail_msg:
|
||||
# Try to capture "max_position_embeddings=131072"
|
||||
m_pos = re.search(r"max_position_embeddings=([\d\.]+)", fail_msg)
|
||||
if m_pos:
|
||||
limit = int(float(m_pos.group(1))) # handle 131072.0
|
||||
log(f" -> Found Model Limit: {limit}")
|
||||
target_len = limit
|
||||
continue
|
||||
|
||||
# Fallback: look for simple parenthesis pattern if the above fails
|
||||
m_derived = re.search(r"derived max_model_len\s*\((\d+)\)", fail_msg)
|
||||
if m_derived:
|
||||
limit = int(m_derived.group(1))
|
||||
log(f" -> Found Model Limit (Legacy): {limit}")
|
||||
target_len = limit
|
||||
continue
|
||||
|
||||
# Case C: Estimated Max Length (New vLLM Safe Limit)
|
||||
# "estimated maximum model length is 111536"
|
||||
m_est = re.search(r"estimated maximum model length is (\d+)", fail_msg)
|
||||
if m_est:
|
||||
limit = int(m_est.group(1))
|
||||
log(f" -> Found vLLM Estimated Limit: {limit}")
|
||||
target_len = limit
|
||||
continue
|
||||
|
||||
# Case D: Generic OOM/Crash
|
||||
target_len = int(target_len * 0.8)
|
||||
log(f" -> Backing off to: {target_len}")
|
||||
|
||||
if target_len < 2048:
|
||||
log(" -> Give up (too small)")
|
||||
break
|
||||
finally:
|
||||
if proc:
|
||||
try: proc.terminate()
|
||||
except: pass
|
||||
try: proc.kill()
|
||||
except: pass
|
||||
proc.wait()
|
||||
force_cleanup()
|
||||
|
||||
return result_data
|
||||
|
||||
def main():
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument("--model", type=str, help="Filter to run only this model (substring match)")
|
||||
parser.add_argument("--steps", type=int, default=-1, help="Number of models to run (default: all)")
|
||||
args = parser.parse_args()
|
||||
|
||||
gpu_count = get_gpu_count()
|
||||
|
||||
# 1. Load existing results to support RESUME
|
||||
results = []
|
||||
if RESULTS_FILE.exists():
|
||||
try:
|
||||
with open(RESULTS_FILE, "r") as f:
|
||||
results = json.load(f)
|
||||
log(f"Loaded {len(results)} previous results. Resuming...")
|
||||
except Exception as e:
|
||||
log(f"Warning: Could not read existing results: {e}")
|
||||
|
||||
count = 0
|
||||
for model in MODELS_TO_RUN:
|
||||
if args.model and args.model not in model:
|
||||
continue
|
||||
|
||||
config = MODEL_TABLE[model]
|
||||
valid_tps = [t for t in config["valid_tp"] if t <= gpu_count]
|
||||
|
||||
for tp in valid_tps:
|
||||
# Track successful seqs for this TP to skip lower utils
|
||||
# effectively: {seqs_count: max_working_util}
|
||||
# Since we iterate high-util -> low-util, if we succeeded already for this 'seqs', we skip.
|
||||
successful_seqs = set()
|
||||
|
||||
# Reset smart limit for each TP (TP2 should not inherit TP1's limit)
|
||||
last_working_len = None
|
||||
|
||||
for util in GPU_UTIL_STEPS:
|
||||
|
||||
for seqs in CONCURRENCY_STEPS:
|
||||
if seqs in successful_seqs:
|
||||
log(f"Skipping {model} (TP={tp}, Util={util}, Seqs={seqs}) - Already succeeded at higher util.")
|
||||
continue
|
||||
|
||||
# Check if we already have this result
|
||||
existing_res = next((r for r in results
|
||||
if r["model"] == model
|
||||
and r["tp"] == tp
|
||||
and str(r["util"]) == str(util)
|
||||
and r["max_seqs"] == seqs), None)
|
||||
|
||||
if existing_res:
|
||||
res = existing_res
|
||||
log(f"Skipping {model} (TP={tp}, Util={util}, Seqs={seqs}) - Found in results.")
|
||||
else:
|
||||
# New run
|
||||
res = run_probe(model, tp, util, seqs, start_limit=last_working_len)
|
||||
results.append(res)
|
||||
|
||||
# Save immediately
|
||||
with open(RESULTS_FILE, "w") as f:
|
||||
json.dump(results, f, indent=2)
|
||||
|
||||
# Update logic for Resume OR New Run:
|
||||
if res["status"] == "success":
|
||||
last_working_len = res["configured_len"]
|
||||
successful_seqs.add(seqs) # Mark this seq count as done for this TP
|
||||
|
||||
# Smart Break: If we failed at this concurrency level (capacity=0),
|
||||
# higher concurrency will also fail.
|
||||
if res["real_capacity"] == 0 or res["status"] == "fail":
|
||||
log(f"Stopping higher concurrency tests for {model} (failed at {seqs} seqs)")
|
||||
break
|
||||
|
||||
count += 1
|
||||
if args.steps != -1 and count >= args.steps and not args.model:
|
||||
break
|
||||
|
||||
# generate_report(results) - Moved to separate script
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -0,0 +1,333 @@
|
||||
[
|
||||
{
|
||||
"model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
|
||||
"tp": 1,
|
||||
"util": "0.95",
|
||||
"max_seqs": 1,
|
||||
"model_limit": 131072,
|
||||
"configured_len": 131072,
|
||||
"real_capacity": 829952,
|
||||
"status": "success",
|
||||
"error": "",
|
||||
"max_context_1_user": 131072
|
||||
},
|
||||
{
|
||||
"model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
|
||||
"tp": 1,
|
||||
"util": "0.95",
|
||||
"max_seqs": 4,
|
||||
"model_limit": 131072,
|
||||
"configured_len": 131072,
|
||||
"real_capacity": 830064,
|
||||
"status": "success",
|
||||
"error": "",
|
||||
"max_context_1_user": 131072
|
||||
},
|
||||
{
|
||||
"model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
|
||||
"tp": 1,
|
||||
"util": "0.95",
|
||||
"max_seqs": 8,
|
||||
"model_limit": 131072,
|
||||
"configured_len": 131072,
|
||||
"real_capacity": 830080,
|
||||
"status": "success",
|
||||
"error": "",
|
||||
"max_context_1_user": 131072
|
||||
},
|
||||
{
|
||||
"model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
|
||||
"tp": 1,
|
||||
"util": "0.95",
|
||||
"max_seqs": 16,
|
||||
"model_limit": 131072,
|
||||
"configured_len": 131072,
|
||||
"real_capacity": 830064,
|
||||
"status": "success",
|
||||
"error": "",
|
||||
"max_context_1_user": 131072
|
||||
},
|
||||
{
|
||||
"model": "google/gemma-3-12b-it",
|
||||
"tp": 1,
|
||||
"util": "0.95",
|
||||
"max_seqs": 1,
|
||||
"model_limit": 131072,
|
||||
"configured_len": 131072,
|
||||
"real_capacity": 246032,
|
||||
"status": "success",
|
||||
"error": "",
|
||||
"max_context_1_user": 131072
|
||||
},
|
||||
{
|
||||
"model": "google/gemma-3-12b-it",
|
||||
"tp": 1,
|
||||
"util": "0.95",
|
||||
"max_seqs": 4,
|
||||
"model_limit": 131072,
|
||||
"configured_len": 131072,
|
||||
"real_capacity": 246064,
|
||||
"status": "success",
|
||||
"error": "",
|
||||
"max_context_1_user": 131072
|
||||
},
|
||||
{
|
||||
"model": "google/gemma-3-12b-it",
|
||||
"tp": 1,
|
||||
"util": "0.95",
|
||||
"max_seqs": 8,
|
||||
"model_limit": 131072,
|
||||
"configured_len": 131072,
|
||||
"real_capacity": 246064,
|
||||
"status": "success",
|
||||
"error": "",
|
||||
"max_context_1_user": 131072
|
||||
},
|
||||
{
|
||||
"model": "google/gemma-3-12b-it",
|
||||
"tp": 1,
|
||||
"util": "0.95",
|
||||
"max_seqs": 16,
|
||||
"model_limit": 131072,
|
||||
"configured_len": 131072,
|
||||
"real_capacity": 246064,
|
||||
"status": "success",
|
||||
"error": "",
|
||||
"max_context_1_user": 131072
|
||||
},
|
||||
{
|
||||
"model": "Qwen/Qwen3-14B-AWQ",
|
||||
"tp": 1,
|
||||
"util": "0.95",
|
||||
"max_seqs": 1,
|
||||
"model_limit": 40960,
|
||||
"configured_len": 0,
|
||||
"real_capacity": 0,
|
||||
"status": "fail",
|
||||
"error": "Verification Failed"
|
||||
},
|
||||
{
|
||||
"model": "Qwen/Qwen3-14B-AWQ",
|
||||
"tp": 1,
|
||||
"util": "0.90",
|
||||
"max_seqs": 1,
|
||||
"model_limit": 40960,
|
||||
"configured_len": 40960,
|
||||
"real_capacity": 655712,
|
||||
"status": "success",
|
||||
"error": "",
|
||||
"max_context_1_user": 40960
|
||||
},
|
||||
{
|
||||
"model": "Qwen/Qwen3-14B-AWQ",
|
||||
"tp": 1,
|
||||
"util": "0.90",
|
||||
"max_seqs": 4,
|
||||
"model_limit": 40960,
|
||||
"configured_len": 40960,
|
||||
"real_capacity": 655616,
|
||||
"status": "success",
|
||||
"error": "",
|
||||
"max_context_1_user": 40960
|
||||
},
|
||||
{
|
||||
"model": "Qwen/Qwen3-14B-AWQ",
|
||||
"tp": 1,
|
||||
"util": "0.90",
|
||||
"max_seqs": 8,
|
||||
"model_limit": 40960,
|
||||
"configured_len": 40960,
|
||||
"real_capacity": 655600,
|
||||
"status": "success",
|
||||
"error": "",
|
||||
"max_context_1_user": 40960
|
||||
},
|
||||
{
|
||||
"model": "Qwen/Qwen3-14B-AWQ",
|
||||
"tp": 1,
|
||||
"util": "0.90",
|
||||
"max_seqs": 16,
|
||||
"model_limit": 40960,
|
||||
"configured_len": 40960,
|
||||
"real_capacity": 655600,
|
||||
"status": "success",
|
||||
"error": "",
|
||||
"max_context_1_user": 40960
|
||||
},
|
||||
{
|
||||
"model": "openai/gpt-oss-20b",
|
||||
"tp": 1,
|
||||
"util": "0.95",
|
||||
"max_seqs": 1,
|
||||
"model_limit": 131072,
|
||||
"configured_len": 131072,
|
||||
"real_capacity": 2232848,
|
||||
"status": "success",
|
||||
"error": "",
|
||||
"max_context_1_user": 131072
|
||||
},
|
||||
{
|
||||
"model": "openai/gpt-oss-20b",
|
||||
"tp": 1,
|
||||
"util": "0.95",
|
||||
"max_seqs": 4,
|
||||
"model_limit": 131072,
|
||||
"configured_len": 131072,
|
||||
"real_capacity": 2232560,
|
||||
"status": "success",
|
||||
"error": "",
|
||||
"max_context_1_user": 131072
|
||||
},
|
||||
{
|
||||
"model": "openai/gpt-oss-20b",
|
||||
"tp": 1,
|
||||
"util": "0.95",
|
||||
"max_seqs": 8,
|
||||
"model_limit": 131072,
|
||||
"configured_len": 131072,
|
||||
"real_capacity": 2232544,
|
||||
"status": "success",
|
||||
"error": "",
|
||||
"max_context_1_user": 131072
|
||||
},
|
||||
{
|
||||
"model": "openai/gpt-oss-20b",
|
||||
"tp": 1,
|
||||
"util": "0.95",
|
||||
"max_seqs": 16,
|
||||
"model_limit": 131072,
|
||||
"configured_len": 131072,
|
||||
"real_capacity": 2232544,
|
||||
"status": "success",
|
||||
"error": "",
|
||||
"max_context_1_user": 131072
|
||||
},
|
||||
{
|
||||
"model": "openai/gpt-oss-120b",
|
||||
"tp": 1,
|
||||
"util": "0.95",
|
||||
"max_seqs": 1,
|
||||
"model_limit": 131072,
|
||||
"configured_len": 131072,
|
||||
"real_capacity": 711360,
|
||||
"status": "success",
|
||||
"error": "",
|
||||
"max_context_1_user": 131072
|
||||
},
|
||||
{
|
||||
"model": "openai/gpt-oss-120b",
|
||||
"tp": 1,
|
||||
"util": "0.95",
|
||||
"max_seqs": 4,
|
||||
"model_limit": 131072,
|
||||
"configured_len": 131072,
|
||||
"real_capacity": 711168,
|
||||
"status": "success",
|
||||
"error": "",
|
||||
"max_context_1_user": 131072
|
||||
},
|
||||
{
|
||||
"model": "openai/gpt-oss-120b",
|
||||
"tp": 1,
|
||||
"util": "0.95",
|
||||
"max_seqs": 8,
|
||||
"model_limit": 131072,
|
||||
"configured_len": 131072,
|
||||
"real_capacity": 711168,
|
||||
"status": "success",
|
||||
"error": "",
|
||||
"max_context_1_user": 131072
|
||||
},
|
||||
{
|
||||
"model": "openai/gpt-oss-120b",
|
||||
"tp": 1,
|
||||
"util": "0.95",
|
||||
"max_seqs": 16,
|
||||
"model_limit": 131072,
|
||||
"configured_len": 131072,
|
||||
"real_capacity": 711168,
|
||||
"status": "success",
|
||||
"error": "",
|
||||
"max_context_1_user": 131072
|
||||
},
|
||||
{
|
||||
"model": "cpatonn/Qwen3-Coder-30B-A3B-Instruct-GPTQ-4bit",
|
||||
"tp": 1,
|
||||
"util": "0.95",
|
||||
"max_seqs": 1,
|
||||
"model_limit": 262144,
|
||||
"configured_len": 262144,
|
||||
"real_capacity": 1097712,
|
||||
"status": "success",
|
||||
"error": "",
|
||||
"max_context_1_user": 262144
|
||||
},
|
||||
{
|
||||
"model": "cpatonn/Qwen3-Coder-30B-A3B-Instruct-GPTQ-4bit",
|
||||
"tp": 1,
|
||||
"util": "0.95",
|
||||
"max_seqs": 4,
|
||||
"model_limit": 262144,
|
||||
"configured_len": 0,
|
||||
"real_capacity": 0,
|
||||
"status": "fail",
|
||||
"error": "Verification Failed"
|
||||
},
|
||||
{
|
||||
"model": "cpatonn/Qwen3-Coder-30B-A3B-Instruct-GPTQ-4bit",
|
||||
"tp": 1,
|
||||
"util": "0.90",
|
||||
"max_seqs": 4,
|
||||
"model_limit": 262144,
|
||||
"configured_len": 209715,
|
||||
"real_capacity": 1029856,
|
||||
"status": "success",
|
||||
"error": "Process died or timed out",
|
||||
"max_context_1_user": 209715
|
||||
},
|
||||
{
|
||||
"model": "cpatonn/Qwen3-Coder-30B-A3B-Instruct-GPTQ-4bit",
|
||||
"tp": 1,
|
||||
"util": "0.90",
|
||||
"max_seqs": 8,
|
||||
"model_limit": 262144,
|
||||
"configured_len": 0,
|
||||
"real_capacity": 0,
|
||||
"status": "fail",
|
||||
"error": "Verification Failed"
|
||||
},
|
||||
{
|
||||
"model": "dazipe/Qwen3-Next-80B-A3B-Instruct-GPTQ-Int4A16",
|
||||
"tp": 1,
|
||||
"util": "0.95",
|
||||
"max_seqs": 1,
|
||||
"model_limit": 262144,
|
||||
"configured_len": 0,
|
||||
"real_capacity": 0,
|
||||
"status": "fail",
|
||||
"error": "Verification Failed"
|
||||
},
|
||||
{
|
||||
"model": "dazipe/Qwen3-Next-80B-A3B-Instruct-GPTQ-Int4A16",
|
||||
"tp": 1,
|
||||
"util": "0.90",
|
||||
"max_seqs": 1,
|
||||
"model_limit": 262144,
|
||||
"configured_len": 262144,
|
||||
"real_capacity": 696320,
|
||||
"status": "success",
|
||||
"error": "",
|
||||
"max_context_1_user": 262144
|
||||
},
|
||||
{
|
||||
"model": "dazipe/Qwen3-Next-80B-A3B-Instruct-GPTQ-Int4A16",
|
||||
"tp": 1,
|
||||
"util": "0.90",
|
||||
"max_seqs": 4,
|
||||
"model_limit": 262144,
|
||||
"configured_len": 0,
|
||||
"real_capacity": 0,
|
||||
"status": "fail",
|
||||
"error": "Verification Failed"
|
||||
}
|
||||
]
|
||||
@@ -0,0 +1,266 @@
|
||||
#!/usr/bin/env python3
|
||||
import subprocess, time, json, sys, os, requests, argparse
|
||||
from pathlib import Path
|
||||
|
||||
# =========================
|
||||
# ⚙️ GLOBAL SETTINGS
|
||||
# =========================
|
||||
|
||||
# HARDWARE: 1x Strix Halo (128GB, RDNA 3.5)
|
||||
GPU_UTIL = "0.90"
|
||||
# 1. THROUGHPUT CONFIG
|
||||
OFF_NUM_PROMPTS = 200
|
||||
OFF_FORCED_OUTPUT = "512"
|
||||
# Default fallback if not specified in MODEL_TABLE
|
||||
DEFAULT_BATCH_TOKENS = "8192"
|
||||
|
||||
# Fallbacks
|
||||
FALLBACK_INPUT_LEN = 1024
|
||||
FALLBACK_OUTPUT_LEN = 512
|
||||
|
||||
RESULTS_DIR = Path("benchmark_results")
|
||||
RESULTS_DIR.mkdir(exist_ok=True)
|
||||
|
||||
# =========================
|
||||
# 🛠️ MODEL CONFIGURATION 🛠️
|
||||
# =========================
|
||||
|
||||
MODEL_TABLE = {
|
||||
# 1. Llama 3.1 8B Instruct
|
||||
# MAD uses 131k tokens. We scale to 32k for 32GB VRAM safety.
|
||||
"meta-llama/Meta-Llama-3.1-8B-Instruct": {
|
||||
"trust_remote": False,
|
||||
"valid_tp": [1, 2],
|
||||
"max_num_seqs": "64",
|
||||
"max_tokens": "32768"
|
||||
},
|
||||
|
||||
"google/gemma-3-12b-it": {
|
||||
"trust_remote": False,
|
||||
"valid_tp": [1, 2],
|
||||
"max_num_seqs": "64",
|
||||
"max_tokens": "32768"
|
||||
},
|
||||
# 2. GPT-OSS 20B (MXFP4)
|
||||
# MAD Row 0 uses 8192. We match this exactly.
|
||||
"openai/gpt-oss-20b": {
|
||||
"trust_remote": True,
|
||||
"valid_tp": [1, 2],
|
||||
"max_num_seqs": "64",
|
||||
"max_tokens": "8192"
|
||||
},
|
||||
|
||||
"openai/gpt-oss-120b": {
|
||||
"trust_remote": True,
|
||||
"valid_tp": [1],
|
||||
"max_num_seqs": "64",
|
||||
"max_tokens": "8192"
|
||||
},
|
||||
|
||||
|
||||
"Qwen/Qwen3-14B-AWQ": {
|
||||
"trust_remote": True,
|
||||
"valid_tp": [1], # Too big for single GPU
|
||||
"max_num_seqs": "32", # Lower concurrency for safety
|
||||
"max_tokens": "16384", # Lower batch size because Eager mode is CPU intensive
|
||||
"enforce_eager": False,
|
||||
"env": {"VLLM_USE_TRITON_AWQ": "1"} # Fixes "Unsupported Hardware" error
|
||||
},
|
||||
|
||||
# 4. Qwen 30B 4-bit
|
||||
"cpatonn/Qwen3-Coder-30B-A3B-Instruct-GPTQ-4bit": {
|
||||
"trust_remote": True,
|
||||
"enforce_eager": False,
|
||||
"valid_tp": [1, 2],
|
||||
"max_num_seqs": "64",
|
||||
"max_tokens": "32768"
|
||||
},
|
||||
|
||||
# 5. Qwen 80B AWQ (The Big One) [NEW]
|
||||
# Size: ~48GB. Fits on 2x32GB (64GB). Leftover for Cache: ~16GB.
|
||||
# Config: 20k ctx fits in that cache. Eager mode required for stability.
|
||||
"dazipe/Qwen3-Next-80B-A3B-Instruct-GPTQ-Int4A16": {
|
||||
"trust_remote": True,
|
||||
"valid_tp": [1], # Too big for single GPU
|
||||
"max_num_seqs": "32", # Lower concurrency for safety
|
||||
"max_tokens": "16384", # Lower batch size because Eager mode is CPU intensive
|
||||
"enforce_eager": True,
|
||||
"env": {"VLLM_USE_TRITON_AWQ": "1"} # Fixes "Unsupported Hardware" error
|
||||
},
|
||||
|
||||
}
|
||||
|
||||
MODELS_TO_RUN = [
|
||||
#"meta-llama/Meta-Llama-3.1-8B-Instruct",
|
||||
#"google/gemma-3-12b-it",
|
||||
#"Qwen/Qwen3-14B-AWQ",
|
||||
#"openai/gpt-oss-20b",
|
||||
#"openai/gpt-oss-120b",
|
||||
"cpatonn/Qwen3-Coder-30B-A3B-Instruct-GPTQ-4bit",
|
||||
"dazipe/Qwen3-Next-80B-A3B-Instruct-GPTQ-Int4A16",
|
||||
]
|
||||
|
||||
# =========================
|
||||
# UTILS
|
||||
# =========================
|
||||
|
||||
def log(msg): print(f"\n[BENCH] {msg}")
|
||||
|
||||
def get_gpu_count():
|
||||
try:
|
||||
# Using rocm-smi --showid to list GPUs.
|
||||
# Output format: "GPU[0] : Device Name: ..."
|
||||
res = subprocess.run(["rocm-smi", "--showid"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
|
||||
if res.returncode == 0:
|
||||
# Filter specifically for the target GPU as requested
|
||||
# target_gpu = "AMD Radeon AI PRO R9700"
|
||||
# count = 0
|
||||
# for line in res.stdout.strip().split('\n'):
|
||||
# if "Device Name" in line and target_gpu in line:
|
||||
# count += 1
|
||||
|
||||
# return count if count > 0 else 1
|
||||
return 1 # Force return 1 for Strix Halo APU
|
||||
else:
|
||||
log("rocm-smi failed, defaulting to 1 GPU (Hardcoded Fallback)")
|
||||
return 1
|
||||
except Exception as e:
|
||||
log(f"Error detecting GPUs: {e}, defaulting to 1 GPU")
|
||||
return 1
|
||||
|
||||
def kill_vllm():
|
||||
subprocess.run("pgrep -f 'vllm serve' | xargs -r kill -9",
|
||||
shell=True, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
|
||||
time.sleep(5)
|
||||
|
||||
def nuke_vllm_cache():
|
||||
cache = Path.home() / ".cache" / "vllm"
|
||||
if cache.exists():
|
||||
try:
|
||||
subprocess.run(["rm", "-rf", str(cache)], check=True)
|
||||
cache.mkdir(parents=True, exist_ok=True)
|
||||
time.sleep(2)
|
||||
except: pass
|
||||
|
||||
def get_dataset():
|
||||
data_path = Path("ShareGPT_V3_unfiltered_cleaned_split.json")
|
||||
if data_path.exists(): return str(data_path)
|
||||
|
||||
log("Downloading ShareGPT dataset...")
|
||||
url = "https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json"
|
||||
try:
|
||||
r = requests.get(url, stream=True, timeout=15)
|
||||
r.raise_for_status()
|
||||
with open(data_path, 'wb') as f:
|
||||
for chunk in r.iter_content(chunk_size=8192): f.write(chunk)
|
||||
return str(data_path)
|
||||
except Exception as e:
|
||||
log(f"WARNING: ShareGPT download failed ({e}). using RANDOM.")
|
||||
return None
|
||||
|
||||
|
||||
|
||||
def get_model_args(model, tp_size):
|
||||
config = MODEL_TABLE.get(model, {"max_num_seqs": "32"})
|
||||
|
||||
# Allow per-model GPU utilization override
|
||||
util = config.get("gpu_util", GPU_UTIL)
|
||||
|
||||
cmd = [
|
||||
"--model", model,
|
||||
"--gpu-memory-utilization", util,
|
||||
"--dtype", "auto",
|
||||
"--tensor-parallel-size", str(tp_size),
|
||||
"--max-num-seqs", config["max_num_seqs"]
|
||||
]
|
||||
|
||||
# Optional: if a model really needs a hard limit, we can still support "ctx" in config,
|
||||
# but by default we rely on auto.
|
||||
if "ctx" in config:
|
||||
cmd.extend(["--max-model-len", config["ctx"]])
|
||||
|
||||
if config.get("trust_remote"): cmd.append("--trust-remote-code")
|
||||
if config.get("enforce_eager"): cmd.append("--enforce-eager")
|
||||
|
||||
return cmd
|
||||
|
||||
def run_throughput(model, tp_size):
|
||||
if tp_size not in MODEL_TABLE[model]["valid_tp"]: return
|
||||
|
||||
model_safe = model.replace("/", "_")
|
||||
output_file = RESULTS_DIR / f"{model_safe}_tp{tp_size}_throughput.json"
|
||||
|
||||
if output_file.exists():
|
||||
log(f"SKIP Throughput {model} (TP={tp_size})")
|
||||
return
|
||||
|
||||
dataset_path = get_dataset()
|
||||
dataset_args = ["--dataset-name", "sharegpt", "--dataset-path", dataset_path] if dataset_path else ["--input-len", "1024"]
|
||||
|
||||
# Retrieve Model-Specific Batch Tokens
|
||||
batch_tokens = MODEL_TABLE[model].get("max_tokens", DEFAULT_BATCH_TOKENS)
|
||||
|
||||
log(f"START Throughput {model} (TP={tp_size}) [Batch: {batch_tokens}]...")
|
||||
kill_vllm()
|
||||
nuke_vllm_cache()
|
||||
|
||||
cmd = ["vllm", "bench", "throughput"] + get_model_args(model, tp_size)
|
||||
cmd.extend([
|
||||
"--num-prompts", str(OFF_NUM_PROMPTS),
|
||||
"--max-num-batched-tokens", batch_tokens,
|
||||
"--output-len", OFF_FORCED_OUTPUT,
|
||||
"--output-json", str(output_file),
|
||||
"--disable-log-stats"
|
||||
])
|
||||
cmd.extend(dataset_args)
|
||||
|
||||
# ENV Setup: Global + Model Specific
|
||||
env = os.environ.copy()
|
||||
|
||||
# Inject model specific env vars (e.g. for AWQ)
|
||||
model_env = MODEL_TABLE[model].get("env", {})
|
||||
env.update(model_env)
|
||||
|
||||
try:
|
||||
subprocess.run(cmd, check=True, env=env)
|
||||
except:
|
||||
log(f"ERROR: Throughput failed {model}")
|
||||
|
||||
|
||||
|
||||
def print_summary(tps):
|
||||
print(f"\n{'MODEL':<40} | {'TP':<2} | {'TOK/S':<8}")
|
||||
print("-" * 60)
|
||||
|
||||
for m in MODELS_TO_RUN:
|
||||
msafe = m.replace("/", "_")
|
||||
for tp in tps:
|
||||
if tp not in MODEL_TABLE[m]["valid_tp"]: continue
|
||||
|
||||
try:
|
||||
tdata = json.loads((RESULTS_DIR / f"{msafe}_tp{tp}_throughput.json").read_text())
|
||||
tok_s = f"{tdata.get('tokens_per_second', 0):.1f}"
|
||||
except: tok_s = "N/A"
|
||||
|
||||
name_cell = m.split('/')[-1]
|
||||
print(f"{name_cell:<40} | {tp:<2} | {tok_s:<8}")
|
||||
print("-" * 60)
|
||||
|
||||
if __name__ == "__main__":
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument("--tp", type=int, nargs="+", default=[1])
|
||||
args = parser.parse_args()
|
||||
|
||||
gpu_count = get_gpu_count()
|
||||
log(f"Detected {gpu_count} AMD GPU(s)")
|
||||
|
||||
valid_tp_args = [t for t in args.tp if t <= gpu_count]
|
||||
if not valid_tp_args:
|
||||
log(f"Requested TP={args.tp} but only {gpu_count} GPU(s) detected. Nothing to run.")
|
||||
sys.exit(0)
|
||||
|
||||
kill_vllm()
|
||||
for tp in valid_tp_args:
|
||||
for m in MODELS_TO_RUN:
|
||||
run_throughput(m, tp)
|
||||
print_summary(valid_tp_args)
|
||||
@@ -0,0 +1,401 @@
|
||||
:root {
|
||||
--bg: #f8fafc;
|
||||
--surface: #ffffff;
|
||||
--text: #0f172a;
|
||||
--text-sub: #64748b;
|
||||
--accent: #d90007;
|
||||
/* AMD Red */
|
||||
--accent-fade: #fff0f0;
|
||||
--border: #e2e8f0;
|
||||
--font: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, sans-serif;
|
||||
}
|
||||
|
||||
body {
|
||||
margin: 0;
|
||||
background: var(--bg);
|
||||
color: var(--text);
|
||||
font-family: var(--font);
|
||||
display: flex;
|
||||
flex-direction: column;
|
||||
height: 100vh;
|
||||
overflow-y: auto;
|
||||
}
|
||||
|
||||
header,
|
||||
.controls,
|
||||
.panel-split,
|
||||
#tables {
|
||||
max-width: 860px;
|
||||
margin: 0 auto;
|
||||
width: 100%;
|
||||
box-sizing: border-box;
|
||||
}
|
||||
|
||||
table {
|
||||
border-collapse: collapse;
|
||||
width: 100%;
|
||||
table-layout: fixed;
|
||||
}
|
||||
|
||||
|
||||
h1 {
|
||||
margin: 0;
|
||||
font-size: 1.25rem;
|
||||
font-weight: 600;
|
||||
}
|
||||
|
||||
p {
|
||||
margin: 4px 0 0;
|
||||
font-size: 0.875rem;
|
||||
color: var(--text-sub);
|
||||
}
|
||||
|
||||
.legend {
|
||||
margin-top: 12px;
|
||||
display: flex;
|
||||
align-items: center;
|
||||
gap: 12px;
|
||||
}
|
||||
|
||||
.legend label {
|
||||
font-size: 0.75rem;
|
||||
font-weight: 600;
|
||||
text-transform: uppercase;
|
||||
color: var(--text-sub);
|
||||
}
|
||||
|
||||
.legend-pills {
|
||||
display: flex;
|
||||
gap: 8px;
|
||||
}
|
||||
|
||||
.legend-pill {
|
||||
cursor: default !important;
|
||||
}
|
||||
|
||||
.legend-pill-default::before {
|
||||
content: "";
|
||||
display: inline-block;
|
||||
width: 8px;
|
||||
height: 8px;
|
||||
background: #cbd5e1;
|
||||
border-radius: 50%;
|
||||
margin-right: 6px;
|
||||
}
|
||||
|
||||
.legend-pill-dual::before {
|
||||
content: "";
|
||||
display: inline-block;
|
||||
width: 8px;
|
||||
height: 8px;
|
||||
background: #d90007;
|
||||
border-radius: 50%;
|
||||
margin-right: 6px;
|
||||
}
|
||||
|
||||
.controls {
|
||||
background: var(--surface);
|
||||
border-bottom: 1px solid var(--border);
|
||||
padding: 12px 24px;
|
||||
display: flex;
|
||||
gap: 24px;
|
||||
align-items: center;
|
||||
flex-shrink: 0;
|
||||
}
|
||||
|
||||
.control {
|
||||
display: flex;
|
||||
flex-direction: column;
|
||||
gap: 4px;
|
||||
}
|
||||
|
||||
.control label {
|
||||
font-size: 0.75rem;
|
||||
font-weight: 600;
|
||||
color: var(--text-sub);
|
||||
}
|
||||
|
||||
input[type="text"],
|
||||
select {
|
||||
padding: 6px 10px;
|
||||
border: 1px solid var(--border);
|
||||
border-radius: 4px;
|
||||
font-size: 0.875rem;
|
||||
background: var(--bg);
|
||||
min-width: 180px;
|
||||
}
|
||||
|
||||
.range-wrap {
|
||||
position: relative;
|
||||
width: 200px;
|
||||
height: 20px;
|
||||
}
|
||||
|
||||
.range-track {
|
||||
position: absolute;
|
||||
top: 50%;
|
||||
left: 0;
|
||||
right: 0;
|
||||
height: 4px;
|
||||
background: #e3e7f1;
|
||||
border-radius: 2px;
|
||||
transform: translateY(-50%);
|
||||
}
|
||||
|
||||
input[type=range] {
|
||||
position: absolute;
|
||||
width: 100%;
|
||||
pointer-events: none;
|
||||
appearance: none;
|
||||
background: none;
|
||||
margin: 0;
|
||||
top: 50%;
|
||||
transform: translateY(-50%);
|
||||
}
|
||||
|
||||
input[type=range]::-webkit-slider-thumb {
|
||||
pointer-events: auto;
|
||||
appearance: none;
|
||||
width: 16px;
|
||||
height: 16px;
|
||||
border-radius: 50%;
|
||||
background: var(--surface);
|
||||
border: 2px solid var(--accent);
|
||||
cursor: pointer;
|
||||
}
|
||||
|
||||
.range-values {
|
||||
font-size: 0.75rem;
|
||||
color: var(--text-sub);
|
||||
margin-top: 4px;
|
||||
text-align: center;
|
||||
}
|
||||
|
||||
.panel {
|
||||
flex: 1;
|
||||
display: flex;
|
||||
flex-direction: column;
|
||||
overflow: hidden;
|
||||
}
|
||||
|
||||
.panel.compact {
|
||||
flex: 0 0 auto;
|
||||
}
|
||||
|
||||
#tables-panel {
|
||||
flex: 1;
|
||||
background: var(--bg);
|
||||
padding: 0;
|
||||
overflow-y: auto;
|
||||
margin-top: 24px;
|
||||
margin-bottom: 40px;
|
||||
}
|
||||
|
||||
.panel-split {
|
||||
display: flex;
|
||||
justify-content: space-between;
|
||||
align-items: center;
|
||||
padding: 12px 24px;
|
||||
background: var(--surface);
|
||||
border-bottom: 1px solid var(--border);
|
||||
}
|
||||
|
||||
.backend-header {
|
||||
display: flex;
|
||||
flex-direction: column;
|
||||
gap: 8px;
|
||||
}
|
||||
|
||||
.backend-label {
|
||||
display: flex;
|
||||
align-items: center;
|
||||
gap: 12px;
|
||||
}
|
||||
|
||||
.backend-label label {
|
||||
font-size: 0.75rem;
|
||||
font-weight: 600;
|
||||
color: var(--text-sub);
|
||||
text-transform: uppercase;
|
||||
}
|
||||
|
||||
.backend-list {
|
||||
display: flex;
|
||||
gap: 16px;
|
||||
flex-wrap: wrap;
|
||||
}
|
||||
|
||||
.backend-item {
|
||||
display: flex;
|
||||
align-items: center;
|
||||
gap: 6px;
|
||||
font-size: 0.875rem;
|
||||
cursor: pointer;
|
||||
user-select: none;
|
||||
}
|
||||
|
||||
.test-block {
|
||||
margin-bottom: 32px;
|
||||
background: var(--surface);
|
||||
border-top: 1px solid var(--border);
|
||||
border-bottom: 1px solid var(--border);
|
||||
}
|
||||
|
||||
h2 {
|
||||
padding: 16px 24px;
|
||||
margin: 0;
|
||||
font-size: 1rem;
|
||||
background: #f1f5f9;
|
||||
color: var(--text);
|
||||
border-bottom: 1px solid var(--border);
|
||||
}
|
||||
|
||||
.table-wrap {
|
||||
position: relative;
|
||||
overflow: hidden;
|
||||
}
|
||||
|
||||
.table-scroll {
|
||||
overflow-x: auto;
|
||||
padding-bottom: 12px;
|
||||
/* Scrollbar space */
|
||||
}
|
||||
|
||||
/* ... */
|
||||
|
||||
.best {
|
||||
background: #f0fdf4;
|
||||
}
|
||||
|
||||
.cell-error {
|
||||
color: #ef4444;
|
||||
font-size: 0.75rem;
|
||||
}
|
||||
|
||||
.cell-empty {
|
||||
color: var(--border);
|
||||
font-size: 0.75rem;
|
||||
font-style: italic;
|
||||
}
|
||||
|
||||
/* Resize Overlay */
|
||||
.resize-overlay {
|
||||
position: absolute;
|
||||
top: 0;
|
||||
left: 0;
|
||||
pointer-events: none;
|
||||
z-index: 5;
|
||||
}
|
||||
|
||||
.resize-bar {
|
||||
position: absolute;
|
||||
top: 0;
|
||||
width: 6px;
|
||||
height: 100%;
|
||||
cursor: col-resize;
|
||||
pointer-events: auto;
|
||||
/* invisible usually, but can hover */
|
||||
}
|
||||
|
||||
.resize-bar:hover {
|
||||
background: rgba(0, 0, 0, 0.05);
|
||||
}
|
||||
|
||||
.resize-handle {
|
||||
position: absolute;
|
||||
right: 0;
|
||||
top: 0;
|
||||
bottom: 0;
|
||||
width: 4px;
|
||||
cursor: col-resize;
|
||||
}
|
||||
|
||||
.backend-header.dragging {
|
||||
opacity: 0.5;
|
||||
}
|
||||
|
||||
.backend-header.drop-target {
|
||||
border-left: 2px solid var(--accent);
|
||||
}
|
||||
|
||||
::-webkit-scrollbar {
|
||||
width: 8px;
|
||||
height: 8px;
|
||||
}
|
||||
|
||||
::-webkit-scrollbar-track {
|
||||
background: transparent;
|
||||
}
|
||||
|
||||
::-webkit-scrollbar-thumb {
|
||||
background: #cbd5e1;
|
||||
border-radius: 4px;
|
||||
}
|
||||
|
||||
/* Modal Styles */
|
||||
.modal {
|
||||
position: fixed;
|
||||
top: 0;
|
||||
left: 0;
|
||||
width: 100vw;
|
||||
height: 100vh;
|
||||
background: rgba(0, 0, 0, 0.4);
|
||||
display: flex;
|
||||
justify-content: center;
|
||||
align-items: center;
|
||||
z-index: 1000;
|
||||
opacity: 1;
|
||||
transition: opacity 0.2s;
|
||||
}
|
||||
|
||||
.modal.hidden {
|
||||
opacity: 0;
|
||||
pointer-events: none;
|
||||
}
|
||||
|
||||
.modal-content {
|
||||
background: var(--surface);
|
||||
padding: 24px 32px;
|
||||
border-radius: 8px;
|
||||
box-shadow: 0 4px 20px rgba(0, 0, 0, 0.15);
|
||||
max-width: 500px;
|
||||
width: 90%;
|
||||
position: relative;
|
||||
transform: translateY(0);
|
||||
transition: transform 0.2s;
|
||||
}
|
||||
|
||||
.modal.hidden .modal-content {
|
||||
transform: translateY(20px);
|
||||
}
|
||||
|
||||
.modal-close {
|
||||
position: absolute;
|
||||
top: 12px;
|
||||
right: 12px;
|
||||
background: transparent;
|
||||
border: none;
|
||||
font-size: 1.5rem;
|
||||
line-height: 1;
|
||||
color: var(--text-sub);
|
||||
cursor: pointer;
|
||||
padding: 4px;
|
||||
}
|
||||
|
||||
.modal-close:hover {
|
||||
color: var(--text);
|
||||
}
|
||||
|
||||
.modal-content h2 {
|
||||
margin-top: 0;
|
||||
background: none;
|
||||
border: none;
|
||||
padding: 0;
|
||||
font-size: 1.25rem;
|
||||
margin-bottom: 12px;
|
||||
}
|
||||
|
||||
.modal-content p {
|
||||
margin-bottom: 12px;
|
||||
line-height: 1.5;
|
||||
}
|
||||
@@ -0,0 +1,542 @@
|
||||
const K_SIGMA = 1.0;
|
||||
const MIN_TOL = 0.25;
|
||||
const MODEL_COL_WIDTH = 300;
|
||||
// Winner column removed
|
||||
|
||||
const state = {
|
||||
envs: ["TP1", "TP2"],
|
||||
backendOrder: ["TP1", "TP2"],
|
||||
columnWidths: { "TP1": 260, "TP2": 260 },
|
||||
filters: {
|
||||
search: "",
|
||||
quant: "",
|
||||
backends: new Set(["TP1", "TP2"]),
|
||||
sizeLo: null,
|
||||
sizeHi: null,
|
||||
},
|
||||
ui: {},
|
||||
sizeStats: { min: Infinity, max: -Infinity },
|
||||
draggingEnv: null,
|
||||
quantOptions: [],
|
||||
};
|
||||
|
||||
document.addEventListener("DOMContentLoaded", async () => {
|
||||
cacheUI();
|
||||
setupModals();
|
||||
try {
|
||||
const res = await fetch("results.json");
|
||||
const data = await res.json();
|
||||
prepareData(data?.runs || []);
|
||||
initializeControls();
|
||||
renderTables();
|
||||
} catch (err) {
|
||||
console.error("Failed to load results.json", err);
|
||||
state.ui.stats.textContent = "Failed to load results.json";
|
||||
}
|
||||
});
|
||||
|
||||
function cacheUI() {
|
||||
state.ui = {
|
||||
search: document.getElementById("filter-search"),
|
||||
quant: document.getElementById("filter-quant"),
|
||||
backendList: document.getElementById("backend-list"),
|
||||
backendAll: document.getElementById("backend-all"),
|
||||
backendNone: document.getElementById("backend-none"),
|
||||
sizeLo: document.getElementById("sizeLo"),
|
||||
sizeHi: document.getElementById("sizeHi"),
|
||||
sizeTrack: document.getElementById("sizeTrack"),
|
||||
sizeLoVal: document.getElementById("sizeLoVal"),
|
||||
sizeHiVal: document.getElementById("sizeHiVal"),
|
||||
stats: document.getElementById("stats-line"),
|
||||
resetBtn: document.getElementById("reset-layout"),
|
||||
tables: document.getElementById("tables"),
|
||||
// Modal hooks
|
||||
tp1ModalOpen: document.getElementById("tp1-modal-open"),
|
||||
tp2ModalOpen: document.getElementById("tp2-modal-open"),
|
||||
tp1Modal: document.getElementById("tp1-modal"),
|
||||
tp2Modal: document.getElementById("tp2-modal"),
|
||||
tp1ModalClose: document.getElementById("tp1-modal-close"),
|
||||
tp2ModalClose: document.getElementById("tp2-modal-close"),
|
||||
};
|
||||
}
|
||||
|
||||
function setupModals() {
|
||||
const modalConfigs = [
|
||||
{ open: state.ui.tp1ModalOpen, modal: state.ui.tp1Modal, close: state.ui.tp1ModalClose },
|
||||
{ open: state.ui.tp2ModalOpen, modal: state.ui.tp2Modal, close: state.ui.tp2ModalClose },
|
||||
];
|
||||
|
||||
modalConfigs.forEach(({ open, modal, close }) => {
|
||||
if (!open || !modal) return;
|
||||
const openModal = () => modal.classList.remove("hidden");
|
||||
const closeModal = () => modal.classList.add("hidden");
|
||||
open.addEventListener("click", openModal);
|
||||
close?.addEventListener("click", closeModal);
|
||||
modal.addEventListener("click", (e) => {
|
||||
if (e.target === modal) closeModal();
|
||||
});
|
||||
document.addEventListener("keydown", (e) => {
|
||||
if (e.key === "Escape" && !modal.classList.contains("hidden")) {
|
||||
closeModal();
|
||||
}
|
||||
});
|
||||
});
|
||||
}
|
||||
|
||||
function prepareData(runs) {
|
||||
const quantSet = new Set();
|
||||
// Tests map: TestName -> { name: ..., models: Map(ModelName -> Row) }
|
||||
const testsMap = new Map();
|
||||
|
||||
for (const run of runs) {
|
||||
if (!run.test) continue;
|
||||
const testKey = run.test;
|
||||
|
||||
if (run.quant) quantSet.add(run.quant.toUpperCase());
|
||||
|
||||
if (!testsMap.has(testKey)) {
|
||||
testsMap.set(testKey, { name: testKey, models: new Map() });
|
||||
}
|
||||
const testEntry = testsMap.get(testKey);
|
||||
|
||||
const modelName = run.model_clean || run.model;
|
||||
|
||||
if (!testEntry.models.has(modelName)) {
|
||||
testEntry.models.set(modelName, {
|
||||
model: modelName,
|
||||
quant: (run.quant || "Unknown").toUpperCase(),
|
||||
sizeB: run.name_params_b ?? run.params_b ?? null,
|
||||
backends: {},
|
||||
search_blob: [modelName, run.quant, run.env, run.test]
|
||||
.filter(Boolean)
|
||||
.map((s) => s.toString().toLowerCase())
|
||||
.join(" "),
|
||||
});
|
||||
}
|
||||
|
||||
const row = testEntry.models.get(modelName);
|
||||
|
||||
// Update stats
|
||||
if (row.sizeB != null) {
|
||||
state.sizeStats.min = Math.min(state.sizeStats.min, row.sizeB);
|
||||
state.sizeStats.max = Math.max(state.sizeStats.max, row.sizeB);
|
||||
}
|
||||
|
||||
// Add backend data
|
||||
// run.env comes from python script as "TP1" or "TP2"
|
||||
const env = run.env;
|
||||
row.backends[env] = {
|
||||
mean: typeof run.tps_mean === "number" ? run.tps_mean : null,
|
||||
std: 0, // Not currently parsed
|
||||
error: Boolean(run.error),
|
||||
error_type: run.error_type || null,
|
||||
};
|
||||
}
|
||||
|
||||
state.tests = [...testsMap.values()].sort((a, b) => a.name.localeCompare(b.name));
|
||||
state.quantOptions = [...quantSet].sort();
|
||||
}
|
||||
|
||||
function initializeControls() {
|
||||
const { quant, backendList, search, resetBtn, sizeLo, sizeHi } = state.ui;
|
||||
|
||||
quant.innerHTML = "";
|
||||
const anyOpt = document.createElement("option");
|
||||
anyOpt.value = "";
|
||||
anyOpt.textContent = "Any";
|
||||
quant.appendChild(anyOpt);
|
||||
state.quantOptions.forEach((q) => {
|
||||
const opt = document.createElement("option");
|
||||
opt.value = q;
|
||||
opt.textContent = q;
|
||||
quant.appendChild(opt);
|
||||
});
|
||||
|
||||
renderBackendList();
|
||||
setupSizeSlider();
|
||||
|
||||
search.addEventListener("input", (e) => {
|
||||
state.filters.search = (e.target.value || "").trim().toLowerCase();
|
||||
renderTables();
|
||||
});
|
||||
|
||||
quant.addEventListener("change", (e) => {
|
||||
state.filters.quant = e.target.value;
|
||||
renderTables();
|
||||
});
|
||||
|
||||
backendList.addEventListener("change", (e) => {
|
||||
const checkbox = e.target.closest("input[data-env]");
|
||||
if (!checkbox) return;
|
||||
const env = checkbox.dataset.env;
|
||||
if (checkbox.checked) {
|
||||
state.filters.backends.add(env);
|
||||
} else {
|
||||
state.filters.backends.delete(env);
|
||||
}
|
||||
renderTables();
|
||||
});
|
||||
|
||||
state.ui.backendAll.addEventListener("click", () => {
|
||||
state.filters.backends = new Set(state.envs);
|
||||
renderBackendList();
|
||||
renderTables();
|
||||
});
|
||||
|
||||
state.ui.backendNone.addEventListener("click", () => {
|
||||
state.filters.backends = new Set();
|
||||
renderBackendList();
|
||||
renderTables();
|
||||
});
|
||||
|
||||
sizeLo.addEventListener("input", () => updateSizeUI(true));
|
||||
sizeHi.addEventListener("input", () => updateSizeUI(true));
|
||||
|
||||
resetBtn.addEventListener("click", () => {
|
||||
state.filters.search = "";
|
||||
state.filters.quant = "";
|
||||
state.filters.backends = new Set(state.envs);
|
||||
search.value = "";
|
||||
quant.value = "";
|
||||
renderBackendList();
|
||||
setupSizeSlider();
|
||||
renderTables();
|
||||
});
|
||||
}
|
||||
|
||||
function renderBackendList() {
|
||||
const container = state.ui.backendList;
|
||||
container.innerHTML = "";
|
||||
state.backendOrder.forEach((env) => {
|
||||
const label = document.createElement("label");
|
||||
label.className = "backend-item";
|
||||
const checkbox = document.createElement("input");
|
||||
checkbox.type = "checkbox";
|
||||
checkbox.dataset.env = env;
|
||||
checkbox.checked = state.filters.backends.has(env);
|
||||
label.appendChild(checkbox);
|
||||
|
||||
const baseSpan = document.createElement("span");
|
||||
baseSpan.textContent = env;
|
||||
label.appendChild(baseSpan);
|
||||
|
||||
container.appendChild(label);
|
||||
});
|
||||
}
|
||||
|
||||
function setupSizeSlider() {
|
||||
const { sizeLo, sizeHi } = state.ui;
|
||||
const minRaw = state.sizeStats.min === Infinity ? 0 : Math.floor(state.sizeStats.min || 0);
|
||||
const maxRaw = state.sizeStats.max === -Infinity ? 0 : Math.ceil(state.sizeStats.max || 0);
|
||||
const minB = Math.max(0, minRaw);
|
||||
const maxB = Math.max(minB, maxRaw);
|
||||
|
||||
[sizeLo, sizeHi].forEach((inp) => {
|
||||
inp.min = minB;
|
||||
inp.max = maxB;
|
||||
inp.step = 1;
|
||||
});
|
||||
|
||||
sizeLo.value = minB;
|
||||
sizeHi.value = maxB;
|
||||
sizeLo.style.zIndex = 2;
|
||||
sizeHi.style.zIndex = 1;
|
||||
updateSizeUI(false);
|
||||
}
|
||||
|
||||
function updateSizeUI(triggerRender) {
|
||||
const { sizeLo, sizeHi, sizeLoVal, sizeHiVal, sizeTrack } = state.ui;
|
||||
if (+sizeLo.value > +sizeHi.value) {
|
||||
if (document.activeElement === sizeLo) {
|
||||
sizeHi.value = sizeLo.value;
|
||||
} else {
|
||||
sizeLo.value = sizeHi.value;
|
||||
}
|
||||
}
|
||||
sizeLo.style.zIndex = +sizeLo.value >= +sizeHi.max - 1 ? 4 : 2;
|
||||
sizeHi.style.zIndex = +sizeHi.value <= +sizeLo.min + 1 ? 3 : 1;
|
||||
state.filters.sizeLo = +sizeLo.value;
|
||||
state.filters.sizeHi = +sizeHi.value;
|
||||
sizeLoVal.textContent = formatSizeLabel(state.filters.sizeLo);
|
||||
sizeHiVal.textContent = formatSizeLabel(state.filters.sizeHi);
|
||||
const range = (sizeHi.max - sizeLo.min) || 1;
|
||||
const minB = +sizeLo.min;
|
||||
const start = ((state.filters.sizeLo - minB) / range) * 100;
|
||||
const end = ((state.filters.sizeHi - minB) / range) * 100;
|
||||
sizeTrack.style.background = `linear-gradient(to right, #e3e7f1 ${start}%, var(--accent) ${start}%, var(--accent) ${end}%, #e3e7f1 ${end}%)`;
|
||||
if (triggerRender) renderTables();
|
||||
}
|
||||
|
||||
function renderTables() {
|
||||
const backendList = state.backendOrder.filter((env) => state.filters.backends.has(env));
|
||||
const frag = document.createDocumentFragment();
|
||||
let totalRows = 0;
|
||||
|
||||
for (const test of state.tests) {
|
||||
const models = filterModels(test.models);
|
||||
if (!models.length) continue;
|
||||
totalRows += models.length;
|
||||
|
||||
const block = document.createElement("div");
|
||||
block.className = "test-block";
|
||||
const heading = document.createElement("h2");
|
||||
heading.textContent = test.name;
|
||||
block.appendChild(heading);
|
||||
|
||||
const tableWrap = document.createElement("div");
|
||||
tableWrap.className = "table-wrap";
|
||||
const scroller = document.createElement("div");
|
||||
scroller.className = "table-scroll";
|
||||
|
||||
const table = buildSingleTable(models, backendList);
|
||||
scroller.appendChild(table);
|
||||
tableWrap.appendChild(scroller);
|
||||
block.appendChild(tableWrap);
|
||||
setupResizeOverlay(scroller, backendList, table);
|
||||
frag.appendChild(block);
|
||||
}
|
||||
|
||||
state.ui.tables.innerHTML = "";
|
||||
if (frag.childNodes.length) {
|
||||
state.ui.tables.appendChild(frag);
|
||||
} else {
|
||||
state.ui.tables.innerHTML = "<p>No models match the current filters.</p>";
|
||||
}
|
||||
state.ui.stats.textContent = `Showing ${totalRows.toLocaleString()} model rows across ${backendList.length} configurations`;
|
||||
}
|
||||
|
||||
function buildSingleTable(models, backendList) {
|
||||
const table = document.createElement("table");
|
||||
const colgroup = document.createElement("colgroup");
|
||||
const colModel = document.createElement("col");
|
||||
colModel.style.width = `${MODEL_COL_WIDTH}px`;
|
||||
colgroup.appendChild(colModel);
|
||||
// Winner colGroup removed
|
||||
|
||||
backendList.forEach((env) => {
|
||||
const col = document.createElement("col");
|
||||
col.style.width = `${state.columnWidths[env] || 200}px`;
|
||||
col.dataset.env = env;
|
||||
colgroup.appendChild(col);
|
||||
});
|
||||
table.appendChild(colgroup);
|
||||
|
||||
const thead = document.createElement("thead");
|
||||
const headRow = document.createElement("tr");
|
||||
headRow.appendChild(makeHeaderCell("Model", "model"));
|
||||
// Winner header removed
|
||||
|
||||
backendList.forEach((env) => {
|
||||
const th = makeHeaderCell(env, ""); // REMOVED "backend-header" class
|
||||
attachHeaderInteractions(th, env);
|
||||
headRow.appendChild(th);
|
||||
});
|
||||
thead.appendChild(headRow);
|
||||
table.appendChild(thead);
|
||||
|
||||
const tbody = document.createElement("tbody");
|
||||
models.forEach((model) => {
|
||||
const tr = document.createElement("tr");
|
||||
const tdModel = document.createElement("td");
|
||||
tdModel.className = "model";
|
||||
const head = document.createElement("div");
|
||||
head.className = "model-head";
|
||||
const nameSpan = document.createElement("span");
|
||||
nameSpan.className = "model-name";
|
||||
nameSpan.textContent = model.model;
|
||||
head.appendChild(nameSpan);
|
||||
tdModel.appendChild(head);
|
||||
|
||||
const meta = document.createElement("div");
|
||||
meta.className = "meta";
|
||||
meta.textContent = `${model.quant} · ${formatSize(model.sizeB)}`;
|
||||
tdModel.appendChild(meta);
|
||||
tr.appendChild(tdModel);
|
||||
|
||||
// Winner cell removed
|
||||
|
||||
backendList.forEach((env) => {
|
||||
const td = document.createElement("td");
|
||||
td.className = "data-cell";
|
||||
td.dataset.env = env;
|
||||
const cell = model.backends[env];
|
||||
if (!cell) {
|
||||
td.innerHTML = `<span class="cell-empty">N/A</span>`;
|
||||
} else if (cell.error || cell.mean == null) {
|
||||
td.innerHTML = `<span class="cell-error">FAIL</span>`;
|
||||
} else {
|
||||
td.innerHTML = `<div class="measure">${cell.mean.toFixed(2)}</div>`;
|
||||
}
|
||||
tr.appendChild(td);
|
||||
});
|
||||
tbody.appendChild(tr);
|
||||
});
|
||||
table.appendChild(tbody);
|
||||
return table;
|
||||
}
|
||||
|
||||
function makeHeaderCell(label, extra = "") {
|
||||
const th = document.createElement("th");
|
||||
th.textContent = label;
|
||||
if (extra) th.className = extra;
|
||||
return th;
|
||||
}
|
||||
|
||||
function attachHeaderInteractions(th, env) {
|
||||
const width = state.columnWidths[env] || 200;
|
||||
th.style.width = `${width}px`;
|
||||
th.style.minWidth = `${width}px`;
|
||||
th.draggable = true;
|
||||
th.addEventListener("dragstart", (e) => {
|
||||
state.draggingEnv = env;
|
||||
th.classList.add("dragging");
|
||||
e.dataTransfer.effectAllowed = "move";
|
||||
});
|
||||
th.addEventListener("dragend", () => {
|
||||
state.draggingEnv = null;
|
||||
th.classList.remove("dragging");
|
||||
document.querySelectorAll("th.drop-target").forEach((el) => el.classList.remove("drop-target"));
|
||||
});
|
||||
th.addEventListener("dragover", (e) => {
|
||||
if (!state.draggingEnv || state.draggingEnv === env) return;
|
||||
e.preventDefault();
|
||||
th.classList.add("drop-target");
|
||||
});
|
||||
th.addEventListener("dragleave", () => th.classList.remove("drop-target"));
|
||||
th.addEventListener("drop", (e) => {
|
||||
if (!state.draggingEnv || state.draggingEnv === env) return;
|
||||
e.preventDefault();
|
||||
moveBackend(state.draggingEnv, env);
|
||||
th.classList.remove("drop-target");
|
||||
});
|
||||
|
||||
const handle = document.createElement("span");
|
||||
handle.className = "resize-handle";
|
||||
handle.addEventListener("mousedown", (e) => startResize(e, env));
|
||||
th.appendChild(handle);
|
||||
}
|
||||
|
||||
function moveBackend(from, to) {
|
||||
const order = state.backendOrder;
|
||||
const fromIdx = order.indexOf(from);
|
||||
const toIdx = order.indexOf(to);
|
||||
if (fromIdx === -1 || toIdx === -1) return;
|
||||
const [col] = order.splice(fromIdx, 1);
|
||||
order.splice(toIdx, 0, col);
|
||||
renderBackendList();
|
||||
renderTables();
|
||||
}
|
||||
|
||||
function filterModels(modelsMap) {
|
||||
const models = [];
|
||||
for (const model of modelsMap.values()) {
|
||||
if (state.filters.search && !model.search_blob.includes(state.filters.search)) continue;
|
||||
if (state.filters.quant && model.quant !== state.filters.quant) continue;
|
||||
if (model.sizeB != null) {
|
||||
if (state.filters.sizeLo != null && model.sizeB < state.filters.sizeLo - 1e-6) continue;
|
||||
if (state.filters.sizeHi != null && model.sizeB > state.filters.sizeHi + 1e-6) continue;
|
||||
}
|
||||
models.push(model);
|
||||
}
|
||||
models.sort((a, b) => a.model.localeCompare(b.model));
|
||||
return models;
|
||||
}
|
||||
|
||||
function formatSize(size) {
|
||||
if (size == null) return "—";
|
||||
return `${Number(size).toFixed(1)}B`;
|
||||
}
|
||||
|
||||
function formatSizeLabel(size) {
|
||||
if (size >= 1000) return `${(size / 1000).toFixed(1)}kB`;
|
||||
return `${Math.round(size)}B`;
|
||||
}
|
||||
|
||||
function startResize(event, env) {
|
||||
event.preventDefault();
|
||||
event.stopPropagation();
|
||||
const column = state.columnWidths[env] || 200;
|
||||
const startX = event.clientX;
|
||||
const shellRect = state.ui.tables.getBoundingClientRect();
|
||||
const guide = document.createElement("div");
|
||||
guide.className = "resize-line";
|
||||
guide.style.position = "fixed";
|
||||
guide.style.top = `${shellRect.top}px`;
|
||||
guide.style.bottom = `${window.innerHeight - shellRect.bottom}px`;
|
||||
guide.style.left = `${startX}px`;
|
||||
guide.style.width = "2px";
|
||||
guide.style.background = "var(--accent)";
|
||||
guide.style.zIndex = "10";
|
||||
document.body.appendChild(guide);
|
||||
let nextWidth = column;
|
||||
|
||||
const onMove = (e) => {
|
||||
const delta = e.clientX - startX;
|
||||
nextWidth = Math.max(80, column + delta);
|
||||
guide.style.left = `${e.clientX}px`;
|
||||
};
|
||||
|
||||
const onUp = () => {
|
||||
document.removeEventListener("mousemove", onMove);
|
||||
document.removeEventListener("mouseup", onUp);
|
||||
guide.remove();
|
||||
state.columnWidths[env] = nextWidth;
|
||||
renderTables();
|
||||
};
|
||||
|
||||
document.addEventListener("mousemove", onMove);
|
||||
document.addEventListener("mouseup", onUp);
|
||||
}
|
||||
|
||||
function setupResizeOverlay(tableWrap, backendList, table) {
|
||||
let overlay = tableWrap.querySelector(".resize-overlay");
|
||||
if (!overlay) {
|
||||
overlay = document.createElement("div");
|
||||
overlay.className = "resize-overlay";
|
||||
tableWrap.appendChild(overlay);
|
||||
} else {
|
||||
overlay.innerHTML = "";
|
||||
}
|
||||
|
||||
overlay.style.width = `${tableWrap.clientWidth}px`;
|
||||
overlay.style.height = `${table.offsetHeight}px`;
|
||||
|
||||
const bars = [];
|
||||
let offset = MODEL_COL_WIDTH; // Winner column width removed
|
||||
backendList.forEach((env) => {
|
||||
const width = state.columnWidths[env] || 200;
|
||||
const bar = document.createElement("div");
|
||||
bar.className = "resize-bar";
|
||||
bar.dataset.env = env;
|
||||
bar.addEventListener("mousedown", (e) => startResize(e, env));
|
||||
overlay.appendChild(bar);
|
||||
bars.push({ bar, offset, width, env });
|
||||
offset += width;
|
||||
});
|
||||
|
||||
const positionBars = () => {
|
||||
bars.forEach(({ bar, offset, width }) => {
|
||||
const left = offset + width - 3 - tableWrap.scrollLeft;
|
||||
bar.style.left = `${left}px`;
|
||||
});
|
||||
};
|
||||
positionBars();
|
||||
|
||||
if (tableWrap._overlayScroll) {
|
||||
tableWrap.removeEventListener("scroll", tableWrap._overlayScroll);
|
||||
}
|
||||
const onScroll = () => positionBars();
|
||||
tableWrap.addEventListener("scroll", onScroll);
|
||||
tableWrap._overlayScroll = onScroll;
|
||||
|
||||
if (tableWrap._overlayResize) {
|
||||
tableWrap._overlayResize.disconnect();
|
||||
}
|
||||
const resizeObserver = new ResizeObserver(() => {
|
||||
overlay.style.width = `${tableWrap.clientWidth}px`;
|
||||
overlay.style.height = `${table.offsetHeight}px`;
|
||||
positionBars();
|
||||
});
|
||||
resizeObserver.observe(tableWrap);
|
||||
tableWrap._overlayResize = resizeObserver;
|
||||
}
|
||||
+782
@@ -0,0 +1,782 @@
|
||||
<!DOCTYPE html>
|
||||
<html lang="en">
|
||||
|
||||
<head>
|
||||
<meta charset="UTF-8">
|
||||
<meta name="viewport" content="width=device-width, initial-scale=1.0">
|
||||
<title>AMD Strix Halo (gfx1151) vLLM Benchmarks</title>
|
||||
<style>
|
||||
:root {
|
||||
--bg-body: #f9fafb;
|
||||
--bg-card: #ffffff;
|
||||
--text-main: #111827;
|
||||
--text-muted: #6b7280;
|
||||
--border: #e5e7eb;
|
||||
--primary: #ef4444;
|
||||
/* AMD Red-ish */
|
||||
--primary-bg: #fef2f2;
|
||||
--font-sans: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, "Helvetica Neue", Arial, sans-serif;
|
||||
--font-mono: ui-monospace, SFMono-Regular, Menlo, Monaco, Consolas, monospace;
|
||||
}
|
||||
|
||||
body {
|
||||
background-color: var(--bg-body);
|
||||
color: var(--text-main);
|
||||
font-family: var(--font-sans);
|
||||
margin: 0;
|
||||
padding: 20px;
|
||||
line-height: 1.5;
|
||||
}
|
||||
|
||||
.container {
|
||||
max-width: 1000px;
|
||||
margin: 20px auto;
|
||||
}
|
||||
|
||||
/* Header */
|
||||
header {
|
||||
margin-bottom: 20px;
|
||||
text-align: center;
|
||||
}
|
||||
|
||||
h1 {
|
||||
font-size: 2.25rem;
|
||||
font-weight: 800;
|
||||
margin: 0 0 10px 0;
|
||||
letter-spacing: -0.05rem;
|
||||
}
|
||||
|
||||
p.subtitle {
|
||||
color: var(--text-muted);
|
||||
font-size: 1.1rem;
|
||||
margin: 0;
|
||||
}
|
||||
|
||||
/* Controls */
|
||||
.controls {
|
||||
display: flex;
|
||||
gap: 16px;
|
||||
margin-bottom: 24px;
|
||||
background: var(--bg-card);
|
||||
padding: 16px;
|
||||
border-radius: 12px;
|
||||
box-shadow: 0 1px 3px rgba(0, 0, 0, 0.05);
|
||||
border: 1px solid var(--border);
|
||||
align-items: center;
|
||||
flex-wrap: wrap;
|
||||
}
|
||||
|
||||
input[type="text"],
|
||||
select {
|
||||
padding: 10px 14px;
|
||||
border: 1px solid var(--border);
|
||||
border-radius: 8px;
|
||||
font-size: 0.95rem;
|
||||
outline: none;
|
||||
transition: border-color 0.15s;
|
||||
}
|
||||
|
||||
input[type="text"]:focus,
|
||||
select:focus {
|
||||
border-color: var(--primary);
|
||||
box-shadow: 0 0 0 2px var(--primary-bg);
|
||||
}
|
||||
|
||||
.search {
|
||||
flex: 1;
|
||||
min-width: 200px;
|
||||
}
|
||||
|
||||
/* Section Cards */
|
||||
.section-card {
|
||||
background: var(--bg-card);
|
||||
border-radius: 12px;
|
||||
box-shadow: 0 1px 3px rgba(0, 0, 0, 0.05);
|
||||
border: 1px solid var(--border);
|
||||
margin-bottom: 32px;
|
||||
overflow: hidden;
|
||||
}
|
||||
|
||||
.section-header {
|
||||
padding: 12px 16px;
|
||||
border-bottom: 1px solid var(--border);
|
||||
background: #fcfcfc;
|
||||
display: flex;
|
||||
justify-content: space-between;
|
||||
align-items: center;
|
||||
}
|
||||
|
||||
.section-header h2 {
|
||||
margin: 0;
|
||||
font-size: 1.1rem;
|
||||
font-weight: 600;
|
||||
}
|
||||
|
||||
/* Table */
|
||||
.table-responsive {
|
||||
overflow-x: auto;
|
||||
}
|
||||
|
||||
table {
|
||||
width: 100%;
|
||||
border-collapse: collapse;
|
||||
font-size: 0.95rem;
|
||||
}
|
||||
|
||||
th,
|
||||
td {
|
||||
padding: 8px 12px;
|
||||
text-align: left;
|
||||
border-bottom: 1px solid var(--border);
|
||||
}
|
||||
|
||||
th {
|
||||
background: #f9fafb;
|
||||
color: var(--text-muted);
|
||||
font-weight: 600;
|
||||
font-size: 0.75rem;
|
||||
text-transform: uppercase;
|
||||
letter-spacing: 0.05em;
|
||||
}
|
||||
|
||||
tr:last-child td {
|
||||
border-bottom: none;
|
||||
}
|
||||
|
||||
/* Columns */
|
||||
.col-model {
|
||||
width: auto;
|
||||
}
|
||||
|
||||
.col-data {
|
||||
text-align: right;
|
||||
width: 1%;
|
||||
white-space: nowrap;
|
||||
font-family: var(--font-mono);
|
||||
font-feature-settings: "tnum";
|
||||
font-variant-numeric: tabular-nums;
|
||||
}
|
||||
|
||||
/* Model Cell Styling */
|
||||
.model-cell {
|
||||
display: flex;
|
||||
flex-direction: column;
|
||||
}
|
||||
|
||||
.model-name {
|
||||
font-weight: 600;
|
||||
color: var(--text-main);
|
||||
}
|
||||
|
||||
.model-meta {
|
||||
font-size: 0.8rem;
|
||||
color: var(--text-muted);
|
||||
margin-top: 4px;
|
||||
display: flex;
|
||||
gap: 8px;
|
||||
align-items: center;
|
||||
}
|
||||
|
||||
/* Tags */
|
||||
.tag {
|
||||
display: inline-block;
|
||||
padding: 2px 6px;
|
||||
border-radius: 4px;
|
||||
background: #f3f4f6;
|
||||
color: #4b5563;
|
||||
font-size: 0.7rem;
|
||||
font-weight: 500;
|
||||
}
|
||||
|
||||
/* Data Styling */
|
||||
.val {
|
||||
font-weight: 600;
|
||||
}
|
||||
|
||||
.val-na {
|
||||
color: #d1d5db;
|
||||
font-weight: 400;
|
||||
}
|
||||
|
||||
.highlight {
|
||||
color: var(--primary);
|
||||
}
|
||||
|
||||
/* Modal/Overlay */
|
||||
#loading {
|
||||
text-align: center;
|
||||
padding: 40px;
|
||||
color: var(--text-muted);
|
||||
}
|
||||
|
||||
/* Modal Styles */
|
||||
.modal-overlay {
|
||||
position: fixed;
|
||||
top: 0;
|
||||
left: 0;
|
||||
width: 100%;
|
||||
height: 100%;
|
||||
background: rgba(0, 0, 0, 0.5);
|
||||
display: flex;
|
||||
justify-content: center;
|
||||
align-items: center;
|
||||
z-index: 1000;
|
||||
opacity: 0;
|
||||
pointer-events: none;
|
||||
transition: opacity 0.2s ease;
|
||||
}
|
||||
|
||||
.modal-overlay.active {
|
||||
opacity: 1;
|
||||
pointer-events: auto;
|
||||
}
|
||||
|
||||
.modal {
|
||||
background: var(--bg-card);
|
||||
width: 90%;
|
||||
max-width: 600px;
|
||||
border-radius: 12px;
|
||||
box-shadow: 0 10px 25px rgba(0, 0, 0, 0.1);
|
||||
display: flex;
|
||||
flex-direction: column;
|
||||
max-height: 85vh;
|
||||
overflow: hidden;
|
||||
}
|
||||
|
||||
.modal-header {
|
||||
padding: 20px 24px;
|
||||
border-bottom: 1px solid var(--border);
|
||||
display: flex;
|
||||
justify-content: space-between;
|
||||
align-items: center;
|
||||
background: #f9fafb;
|
||||
}
|
||||
|
||||
.modal-header h3 {
|
||||
margin: 0;
|
||||
font-size: 1.25rem;
|
||||
}
|
||||
|
||||
.modal-close {
|
||||
background: none;
|
||||
border: none;
|
||||
font-size: 1.5rem;
|
||||
cursor: pointer;
|
||||
color: var(--text-muted);
|
||||
line-height: 1;
|
||||
}
|
||||
|
||||
.modal-body {
|
||||
padding: 24px;
|
||||
overflow-y: auto;
|
||||
}
|
||||
|
||||
.modal-section {
|
||||
margin-bottom: 24px;
|
||||
}
|
||||
|
||||
.modal-section h4 {
|
||||
margin: 0 0 8px 0;
|
||||
font-size: 0.9rem;
|
||||
text-transform: uppercase;
|
||||
color: var(--text-muted);
|
||||
letter-spacing: 0.05em;
|
||||
}
|
||||
|
||||
.modal-section p {
|
||||
margin: 0;
|
||||
font-size: 0.95rem;
|
||||
color: var(--text-main);
|
||||
}
|
||||
|
||||
.code-block {
|
||||
background: #f3f4f6;
|
||||
padding: 12px;
|
||||
border-radius: 6px;
|
||||
font-family: var(--font-mono);
|
||||
font-size: 0.85rem;
|
||||
color: #374151;
|
||||
margin-top: 8px;
|
||||
white-space: pre-wrap;
|
||||
}
|
||||
|
||||
/* Help Button */
|
||||
.btn-help {
|
||||
background: none;
|
||||
border: 1px solid var(--border);
|
||||
color: var(--text-muted);
|
||||
width: 24px;
|
||||
height: 24px;
|
||||
border-radius: 50%;
|
||||
display: inline-flex;
|
||||
align-items: center;
|
||||
justify-content: center;
|
||||
font-size: 0.85rem;
|
||||
font-weight: 600;
|
||||
cursor: pointer;
|
||||
margin-left: 10px;
|
||||
transition: all 0.2s;
|
||||
}
|
||||
|
||||
.btn-help:hover {
|
||||
border-color: var(--primary);
|
||||
color: var(--primary);
|
||||
background: var(--primary-bg);
|
||||
}
|
||||
|
||||
.section-title-row {
|
||||
display: flex;
|
||||
align-items: center;
|
||||
}
|
||||
|
||||
.section-desc {
|
||||
color: var(--text-muted);
|
||||
font-size: 0.9rem;
|
||||
font-weight: 400;
|
||||
margin-left: 12px;
|
||||
}
|
||||
|
||||
/* Footer */
|
||||
footer {
|
||||
margin-top: 60px;
|
||||
padding-top: 20px;
|
||||
border-top: 1px solid var(--border);
|
||||
color: var(--text-muted);
|
||||
font-size: 0.85rem;
|
||||
line-height: 1.6;
|
||||
}
|
||||
|
||||
.sys-config {
|
||||
display: flex;
|
||||
flex-direction: column;
|
||||
gap: 8px;
|
||||
margin-top: 12px;
|
||||
max-width: 800px;
|
||||
}
|
||||
|
||||
.sys-item {
|
||||
display: grid;
|
||||
grid-template-columns: 140px 1fr;
|
||||
align-items: baseline;
|
||||
}
|
||||
|
||||
.sys-label {
|
||||
font-weight: 600;
|
||||
font-size: 0.75rem;
|
||||
text-transform: uppercase;
|
||||
letter-spacing: 0.05em;
|
||||
color: #9ca3af;
|
||||
}
|
||||
|
||||
/* Tabs */
|
||||
.tab-nav {
|
||||
display: flex;
|
||||
gap: 8px;
|
||||
margin-bottom: 24px;
|
||||
border-bottom: 1px solid var(--border);
|
||||
padding-bottom: 0px;
|
||||
}
|
||||
|
||||
.tab-btn {
|
||||
background: none;
|
||||
border: none;
|
||||
padding: 12px 20px;
|
||||
font-size: 1rem;
|
||||
font-weight: 500;
|
||||
color: var(--text-muted);
|
||||
cursor: pointer;
|
||||
border-bottom: 2px solid transparent;
|
||||
transition: all 0.2s;
|
||||
}
|
||||
|
||||
.tab-btn:hover {
|
||||
color: var(--text-main);
|
||||
}
|
||||
|
||||
.tab-btn.active {
|
||||
color: var(--primary);
|
||||
border-bottom-color: var(--primary);
|
||||
font-weight: 600;
|
||||
}
|
||||
</style>
|
||||
</head>
|
||||
|
||||
<body>
|
||||
|
||||
<div class="container">
|
||||
<header>
|
||||
<h1>AMD Strix Halo (gfx1151) vLLM Benchmarks</h1>
|
||||
<p style="margin: 4px 0 0 0; font-size: 0.9rem;">
|
||||
<a href="https://github.com/kyuz0/amd-strix-halo-vllm-toolboxes/" target="_blank"
|
||||
style="color: var(--primary); text-decoration: none;">View on GitHub →</a>
|
||||
</p>
|
||||
</header>
|
||||
|
||||
<div class="controls">
|
||||
<input type="text" id="searchInput" class="search" placeholder="Search models (e.g. 'llama', 'fp8')..."
|
||||
autocomplete="off">
|
||||
<select id="quantFilter">
|
||||
<option value="">All Quantizations</option>
|
||||
</select>
|
||||
</div>
|
||||
|
||||
<nav id="tabNav" class="tab-nav">
|
||||
<!-- Dynamic Tabs -->
|
||||
</nav>
|
||||
|
||||
<div id="dashboard">
|
||||
<div id="loading">Loading benchmark results...</div>
|
||||
</div>
|
||||
|
||||
<footer>
|
||||
<div style="font-weight: 600; margin-bottom: 8px;">System Configuration</div>
|
||||
<div class="sys-config">
|
||||
<div class="sys-item">
|
||||
<span class="sys-label">System</span>
|
||||
<span>Framework Desktop · AMD Ryzen AI MAX 395+ · 128GB unified RAM</span>
|
||||
</div>
|
||||
<div class="sys-item">
|
||||
<span class="sys-label">OS/Kernel</span>
|
||||
<span>Fedora 42 · Linux 6.18.0-0.rc6.243.vanilla.fc42.x86_64</span>
|
||||
</div>
|
||||
</div>
|
||||
</footer>
|
||||
</div>
|
||||
|
||||
<!-- Modal Overlay -->
|
||||
<div id="modalOverlay" class="modal-overlay">
|
||||
<div class="modal">
|
||||
<div class="modal-header">
|
||||
<h3 id="modalTitle">Benchmark Info</h3>
|
||||
<button class="modal-close" onclick="closeModal()">×</button>
|
||||
</div>
|
||||
<div class="modal-body" id="modalContent">
|
||||
<!-- Dynamic Content -->
|
||||
</div>
|
||||
</div>
|
||||
</div>
|
||||
|
||||
<script>
|
||||
// State
|
||||
let rawRuns = [];
|
||||
let tests = [];
|
||||
let state = {
|
||||
search: "",
|
||||
quant: "",
|
||||
activeTab: "Throughput"
|
||||
};
|
||||
|
||||
// Benchmark Metadata
|
||||
const BENCHMARK_INFO = {
|
||||
"Throughput": {
|
||||
short: "Maximum raw compute capacity (Tokens/Sec).",
|
||||
desc: "Measures the absolute maximum number of tokens the system can generate per second by fully saturating the GPU compute capability.",
|
||||
usecase: "Demonstrates the raw horsepower and architectural efficiency of the hardware/model combo under Heavy Load. This is the theoretical speed limit of the system.",
|
||||
details: "Command: `vllm bench throughput`\nParams: --num-prompts 100 --output-len 512\nMetric: Tokens per Second (higher is better).",
|
||||
unit: " tok/s"
|
||||
},
|
||||
"TTFT": {
|
||||
short: "Time To First Token (Response Latency).",
|
||||
desc: "The 'Time To First Token' is the delay between sending a request and seeing the first character of the response.",
|
||||
usecase: "<b>Responsiveness</b>. Low TTFT makes the AI feel 'snappy' and instant. High TTFT feels like the AI is ignoring you or lagging. We measure at different QPS loads to ensure the server doesn't 'choke' when busy.",
|
||||
context: "<b>QPS = Queries Per Second (Traffic Load)</b>.<br>• QPS 1.0 = 1 user sending a request every second.<br>• QPS 4.0 = 4 users sending requests every second (Simulates High Load).",
|
||||
details: "Command: `vllm bench serve`\nParams: --random-input-len 1024 --random-output-len 512\nMetric: Milliseconds (lower is better).",
|
||||
unit: " ms"
|
||||
},
|
||||
"TPOT": {
|
||||
short: "Time Per Output Token (Streaming Speed).",
|
||||
desc: "The 'Time Per Output Token' measures how fast the text generates *after* the first token appears.",
|
||||
usecase: "<b>1. Fluidity</b>: Industry standard is <50ms (>20 tok/s) for a 'fluid' feeling. Slower feels laggy.<br><b>2. Bottlenecks</b>: We test at <b>QPS 4.0</b> to find memory bandwidth bottlenecks where the GPU can't keep up with multiple users.",
|
||||
context: "<b>QPS = Queries Per Second (Traffic Load)</b>.<br>• QPS 1.0 = Light Load (Ideal conditions)<br>• QPS 4.0 = Heavy Load (Stress Test)",
|
||||
details: "Command: `vllm bench serve`\nParams: --random-input-len 1024 --random-output-len 512\nMetric: Milliseconds (lower is better).",
|
||||
unit: " ms"
|
||||
}
|
||||
};
|
||||
|
||||
const $ = id => document.getElementById(id);
|
||||
|
||||
async function init() {
|
||||
try {
|
||||
const res = await fetch('results.json');
|
||||
const data = await res.json();
|
||||
rawRuns = data.runs || [];
|
||||
processData();
|
||||
render();
|
||||
populateFilters();
|
||||
} catch (e) {
|
||||
$('loading').textContent = "Error loading results.json: " + e.message;
|
||||
console.error(e);
|
||||
}
|
||||
}
|
||||
|
||||
function processData() {
|
||||
const testGroups = {};
|
||||
|
||||
rawRuns.forEach(run => {
|
||||
if (!run.test) return;
|
||||
if (!testGroups[run.test]) {
|
||||
testGroups[run.test] = {
|
||||
name: run.test,
|
||||
models: {}
|
||||
};
|
||||
}
|
||||
|
||||
// Normalize model name
|
||||
const modelName = run.model_clean || run.model;
|
||||
|
||||
if (!testGroups[run.test].models[modelName]) {
|
||||
testGroups[run.test].models[modelName] = {
|
||||
name: modelName,
|
||||
quant: run.quant,
|
||||
params: run.params_b || run.name_params_b,
|
||||
tp1: null,
|
||||
tp2: null
|
||||
};
|
||||
}
|
||||
|
||||
const m = testGroups[run.test].models[modelName];
|
||||
|
||||
// Assign TP value
|
||||
if (run.env === "TP1") m.tp1 = run.tps_mean;
|
||||
if (run.env === "TP2") m.tp2 = run.tps_mean;
|
||||
});
|
||||
|
||||
// Convert map to array for sorting
|
||||
tests = Object.values(testGroups).map(group => {
|
||||
return {
|
||||
name: group.name,
|
||||
models: Object.values(group.models)
|
||||
};
|
||||
});
|
||||
|
||||
// Sort tests: Throughput first, then others alphabetically
|
||||
tests.sort((a, b) => {
|
||||
if (a.name === "Throughput") return -1;
|
||||
if (b.name === "Throughput") return 1;
|
||||
return a.name.localeCompare(b.name);
|
||||
});
|
||||
|
||||
// Set default tab if not set
|
||||
if (!state.activeTab && tests.length > 0) {
|
||||
state.activeTab = tests[0].name;
|
||||
}
|
||||
}
|
||||
|
||||
function populateFilters() {
|
||||
const quants = new Set(rawRuns.map(r => r.quant).filter(Boolean));
|
||||
const sel = $('quantFilter');
|
||||
[...quants].sort().forEach(q => {
|
||||
const opt = document.createElement('option');
|
||||
opt.value = q;
|
||||
opt.textContent = q;
|
||||
sel.appendChild(opt);
|
||||
});
|
||||
|
||||
$('searchInput').addEventListener('input', e => {
|
||||
state.search = e.target.value.toLowerCase();
|
||||
render();
|
||||
});
|
||||
|
||||
sel.addEventListener('change', e => {
|
||||
state.quant = e.target.value;
|
||||
render();
|
||||
});
|
||||
}
|
||||
|
||||
function getBenchmarkMeta(testName) {
|
||||
if (testName.includes("Throughput")) return BENCHMARK_INFO["Throughput"];
|
||||
if (testName.includes("TTFT")) return BENCHMARK_INFO["TTFT"];
|
||||
if (testName.includes("TPOT")) return BENCHMARK_INFO["TPOT"];
|
||||
return null;
|
||||
}
|
||||
|
||||
function render() {
|
||||
const container = $('dashboard');
|
||||
const tabNav = $('tabNav');
|
||||
|
||||
// Render Tabs
|
||||
tabNav.innerHTML = "";
|
||||
tests.forEach(test => {
|
||||
const btn = document.createElement('button');
|
||||
btn.className = `tab-btn ${test.name === state.activeTab ? 'active' : ''}`;
|
||||
btn.textContent = test.name;
|
||||
btn.onclick = () => {
|
||||
state.activeTab = test.name;
|
||||
render();
|
||||
};
|
||||
tabNav.appendChild(btn);
|
||||
});
|
||||
|
||||
// Ensure active tab exists (if search filtered it out logic?)
|
||||
// Actually tabs are based on 'tests' which is processed from raw data, so they exist regardless of filters unless we want to hide tabs with no results.
|
||||
// For now, let's keep tabs static based on available data types.
|
||||
|
||||
container.innerHTML = "";
|
||||
|
||||
// Find active test
|
||||
const activeTest = tests.find(t => t.name === state.activeTab);
|
||||
|
||||
if (!activeTest) {
|
||||
// If invalid tab (e.g. on first load if default doesn't exist), switch to first
|
||||
if (tests.length > 0) {
|
||||
state.activeTab = tests[0].name;
|
||||
// Re-render immediately
|
||||
setTimeout(render, 0);
|
||||
}
|
||||
container.innerHTML = '<div id="loading">No data available.</div>';
|
||||
return;
|
||||
}
|
||||
|
||||
// Render Active Tab Content
|
||||
const test = activeTest;
|
||||
|
||||
// Filter models within this test
|
||||
const models = test.models.filter(m => {
|
||||
const s = state.search;
|
||||
const matchSearch = !s || m.name.toLowerCase().includes(s);
|
||||
|
||||
const q = state.quant;
|
||||
const matchQuant = !q || m.quant === q;
|
||||
|
||||
return matchSearch && matchQuant;
|
||||
});
|
||||
|
||||
if (models.length === 0) {
|
||||
container.innerHTML = '<div id="loading">No models match current filters in this category.</div>';
|
||||
return;
|
||||
}
|
||||
|
||||
// Sorting models by size (small to large), then name
|
||||
models.sort((a, b) => {
|
||||
const pA = parseFloat(a.params) || 0;
|
||||
const pB = parseFloat(b.params) || 0;
|
||||
if (pA !== pB) return pA - pB;
|
||||
return a.name.localeCompare(b.name);
|
||||
});
|
||||
|
||||
const card = document.createElement('div');
|
||||
card.className = "section-card";
|
||||
|
||||
// Metadata resolution
|
||||
const meta = getBenchmarkMeta(test.name);
|
||||
const shortDesc = meta ? `<span class="section-desc">${meta.short}</span>` : "";
|
||||
const helpBtn = meta ? `<button class="btn-help" onclick="openModal('${test.name}')">?</button>` : "";
|
||||
|
||||
const header = document.createElement('div');
|
||||
header.className = "section-header";
|
||||
header.innerHTML = `
|
||||
<div class="section-title-row">
|
||||
<h2>${test.name}</h2>
|
||||
${helpBtn}
|
||||
</div>
|
||||
${shortDesc}
|
||||
`;
|
||||
card.appendChild(header);
|
||||
|
||||
const tableResp = document.createElement('div');
|
||||
tableResp.className = "table-responsive";
|
||||
|
||||
const table = document.createElement('table');
|
||||
const thead = document.createElement('thead');
|
||||
thead.innerHTML = `
|
||||
<tr>
|
||||
<th class="col-model">Model</th>
|
||||
<th class="col-data">TP1</th>
|
||||
</tr>
|
||||
`;
|
||||
table.appendChild(thead);
|
||||
|
||||
const tbody = document.createElement('tbody');
|
||||
models.forEach(m => {
|
||||
const tr = document.createElement('tr');
|
||||
|
||||
// Meta tags
|
||||
let metaHtml = "";
|
||||
if (m.quant) metaHtml += `<span class="tag">${m.quant}</span>`;
|
||||
if (m.params) metaHtml += `<span class="tag">${m.params}B</span>`;
|
||||
|
||||
// Values
|
||||
// Pass unit from meta
|
||||
const unit = meta ? meta.unit : "";
|
||||
const val1 = formatVal(m.tp1, unit);
|
||||
|
||||
tr.innerHTML = `
|
||||
<td>
|
||||
<div class="model-cell">
|
||||
<a href="https://huggingface.co/${m.name}" target="_blank" class="model-name" style="text-decoration: none; color: inherit; border-bottom: 1px dotted #ccc;">${m.name}</a>
|
||||
<div class="model-meta">${metaHtml}</div>
|
||||
</div>
|
||||
</td>
|
||||
<td class="col-data">${val1}</td>
|
||||
`;
|
||||
tbody.appendChild(tr);
|
||||
});
|
||||
table.appendChild(tbody);
|
||||
tableResp.appendChild(table);
|
||||
card.appendChild(tableResp);
|
||||
|
||||
container.appendChild(card);
|
||||
}
|
||||
|
||||
function formatVal(v, unit) {
|
||||
if (v === null || v === undefined) return '<span class="val-na">N/A</span>';
|
||||
if (v === 0) return '<span class="val-na">FAIL</span>';
|
||||
return `<span class="val">${v.toFixed(2)}<span style="font-size:0.8em; color:#888;">${unit}</span></span>`;
|
||||
}
|
||||
|
||||
// Modal Logic
|
||||
function openModal(testName) {
|
||||
const meta = getBenchmarkMeta(testName);
|
||||
if (!meta) return;
|
||||
|
||||
$('modalTitle').textContent = testName;
|
||||
|
||||
let content = `
|
||||
<div class="modal-section">
|
||||
<h4>What is this?</h4>
|
||||
<p>${meta.desc}</p>
|
||||
</div>
|
||||
<div class="modal-section">
|
||||
<h4>Why it matters?</h4>
|
||||
<p>${meta.usecase}</p>
|
||||
</div>`;
|
||||
|
||||
if (meta.context) {
|
||||
content += `
|
||||
<div class="modal-section">
|
||||
<h4>Terminology</h4>
|
||||
<p>${meta.context}</p>
|
||||
</div>`;
|
||||
}
|
||||
|
||||
content += `
|
||||
<div class="modal-section">
|
||||
<h4>Technical Details</h4>
|
||||
<div class="code-block">${meta.details}</div>
|
||||
</div>
|
||||
`;
|
||||
|
||||
$('modalContent').innerHTML = content;
|
||||
$('modalOverlay').classList.add('active');
|
||||
}
|
||||
|
||||
function closeModal() {
|
||||
$('modalOverlay').classList.remove('active');
|
||||
}
|
||||
|
||||
// Close on click outside
|
||||
$('modalOverlay').addEventListener('click', e => {
|
||||
if (e.target === $('modalOverlay')) closeModal();
|
||||
});
|
||||
|
||||
// Close on Escape
|
||||
document.addEventListener('keydown', e => {
|
||||
if (e.key === "Escape") closeModal();
|
||||
});
|
||||
|
||||
init();
|
||||
</script>
|
||||
|
||||
</body>
|
||||
|
||||
</html>
|
||||
@@ -0,0 +1,181 @@
|
||||
|
||||
import os
|
||||
import json
|
||||
import re
|
||||
from pathlib import Path
|
||||
|
||||
# Config
|
||||
BENCHMARK_DIR = Path("../benchmarks/benchmark_results")
|
||||
OUTPUT_FILE = Path("results.json")
|
||||
|
||||
# Regex to parse model name for quantization and parameters
|
||||
# Examples:
|
||||
# "meta-llama/Meta-Llama-3.1-8B-In
|
||||
# struct"
|
||||
# "cpatonn/Qwen3-Coder-30B-A3B-Instruct-GPTQ-4bit"
|
||||
# "RedHatAI/Llama-3.1-8B-Instruct-FP8-block"
|
||||
PARAMS_REGEX = r"(\d+(?:\.\d+)?)B"
|
||||
QUANT_REGEX = r"(FP8|AWQ|GPTQ|BF16|4bit|Int4)"
|
||||
|
||||
def extract_meta(model_name):
|
||||
# Params
|
||||
params_match = re.search(PARAMS_REGEX, model_name, re.IGNORECASE)
|
||||
params_b = float(params_match.group(1)) if params_match else None
|
||||
|
||||
# Quant
|
||||
quant_match = re.search(QUANT_REGEX, model_name, re.IGNORECASE)
|
||||
quant = quant_match.group(1).upper() if quant_match else "BF16" # Default assumption if no tag? Or unknown.
|
||||
# Refine quant if 4bit
|
||||
if quant == "4BIT" or quant == "INT4":
|
||||
if "GPTQ" in model_name: quant = "GPTQ-4bit"
|
||||
elif "AWQ" in model_name: quant = "AWQ-4bit"
|
||||
else: quant = "4-bit"
|
||||
|
||||
return params_b, quant
|
||||
|
||||
def parse_logs():
|
||||
runs = []
|
||||
|
||||
if not BENCHMARK_DIR.exists():
|
||||
print(f"Error: {BENCHMARK_DIR} does not exist!")
|
||||
return []
|
||||
|
||||
print(f"Scanning {BENCHMARK_DIR}...")
|
||||
|
||||
# Files are flat in the dir: {model_safe}_tp{tp}_{type}.json
|
||||
# or latency: {model_safe}_tp{tp}_qps{q}_latency.json
|
||||
|
||||
# We need to group by (model, tp) to form cohesive records if we want,
|
||||
# BUT the webapp expects a list of "runs".
|
||||
# Looking at the example JSON, each "run" is a single test point (e.g. "pp2048 @ d16384" OR "tg32 @ d16384")
|
||||
# Actually, looking at the provided valid example:
|
||||
# "test": "pp512", "tps_mean": 2708.86 ...
|
||||
|
||||
# Our data:
|
||||
# throughput.json -> tokens_per_second. This is usually "decoding" or a mix?
|
||||
# vLLM bench throughput usually streams tokens.
|
||||
# Let's look at what run_vllm_bench.py produces.
|
||||
# Throughput: --input-len 1024 --output-len 512.
|
||||
# This is effectively a mixed batch.
|
||||
# We'll label it "Throughput (1024/512)" or just "Throughput"
|
||||
|
||||
# Latency: qps-based.
|
||||
|
||||
files = list(BENCHMARK_DIR.glob("*.json"))
|
||||
|
||||
for f in files:
|
||||
fname = f.name
|
||||
try:
|
||||
data = json.loads(f.read_text())
|
||||
except:
|
||||
print(f"Skipping bad JSON: {fname}")
|
||||
continue
|
||||
|
||||
# Infer metadata from filename
|
||||
# Format: {model_safe}_tp{tp}_{suffix}
|
||||
# Suffix can be: "throughput.json" or "qps{q}_latency.json"
|
||||
|
||||
# We need model name. The script replaces / with _ in filenames.
|
||||
# But we verify against the known models list? Or just parse string.
|
||||
# We can reconstruct roughly.
|
||||
|
||||
# Split by "_tp" which is a strong delimiter
|
||||
parts = fname.split("_tp")
|
||||
if len(parts) < 2: continue
|
||||
|
||||
model_part = parts[0]
|
||||
rest = parts[1] # "1_throughput.json" or "2_qps1.0_latency.json"
|
||||
|
||||
# TP
|
||||
tp_match = re.match(r"^(\d+)", rest)
|
||||
if not tp_match: continue
|
||||
tp = int(tp_match.group(1))
|
||||
|
||||
# Env mapping
|
||||
env = f"TP{tp}"
|
||||
|
||||
# Model Name Restoration (best effort or matching)
|
||||
# In the script: model.replace("/", "_")
|
||||
# We can reverse this if we have the list, but for now let's just use the clean string?
|
||||
# The webapp uses "model_clean" and "model".
|
||||
# Let's assume standard "org_model" format -> "org/model"
|
||||
if "_" in model_part:
|
||||
# Heuristic: First _ is likely the slash
|
||||
model_display = model_part.replace("_", "/", 1)
|
||||
else:
|
||||
model_display = model_part
|
||||
|
||||
params_b, quant = extract_meta(model_display)
|
||||
|
||||
base_run = {
|
||||
"model": model_display,
|
||||
"model_clean": model_display,
|
||||
"env": env,
|
||||
"gpu_config": "dual" if tp > 1 else "single",
|
||||
"quant": quant,
|
||||
"params_b": params_b,
|
||||
"name_params_b": params_b,
|
||||
# Defaults
|
||||
"backend": "vLLM",
|
||||
"error": False
|
||||
}
|
||||
|
||||
if "throughput" in fname:
|
||||
# Throughput run
|
||||
# data has "tokens_per_second"
|
||||
tps = data.get("tokens_per_second", 0)
|
||||
|
||||
run = base_run.copy()
|
||||
run["test"] = "Throughput"
|
||||
run["tps_mean"] = tps
|
||||
# If tps is 0 or missing, it might be an error?
|
||||
if tps == 0 and "error" in str(data).lower():
|
||||
run["error"] = True
|
||||
|
||||
runs.append(run)
|
||||
|
||||
elif "latency" in fname:
|
||||
# Latency run
|
||||
# raw_output has strings like "Mean TTFT: 12.3 ms", "Mean TPOT: 45.6 ms"
|
||||
raw = data.get("raw_output", "")
|
||||
qps_match = re.search(r"_qps([\d\.]+)_", fname)
|
||||
qps = qps_match.group(1) if qps_match else "?"
|
||||
|
||||
# Extract metrics
|
||||
ttft = 0.0
|
||||
tpot = 0.0
|
||||
|
||||
ttft_m = re.search(r"(?:Mean TTFT|TTFT).*?([\d\.]+)", raw)
|
||||
if ttft_m: ttft = float(ttft_m.group(1))
|
||||
|
||||
tpot_m = re.search(r"(?:Mean TPOT|TPOT).*?([\d\.]+)", raw)
|
||||
if tpot_m: tpot = float(tpot_m.group(1))
|
||||
|
||||
# We create TWO entries? Or how does the webapp handle multiple metrics?
|
||||
# Example webapp table columns are "Backends" showing ONE value.
|
||||
# But grouping is by "Test".
|
||||
# So we can have a test called "TTFT (QPS 1.0)" and "TPOT (QPS 1.0)"
|
||||
|
||||
# Entry 1: TTFT
|
||||
r1 = base_run.copy()
|
||||
r1["test"] = f"TTFT @ QPS {qps}"
|
||||
r1["tps_mean"] = ttft # Using tps_mean field for the numeric value
|
||||
runs.append(r1)
|
||||
|
||||
# Entry 2: TPOT
|
||||
r2 = base_run.copy()
|
||||
r2["test"] = f"TPOT @ QPS {qps}"
|
||||
r2["tps_mean"] = tpot
|
||||
runs.append(r2)
|
||||
|
||||
return runs
|
||||
|
||||
if __name__ == "__main__":
|
||||
data = {"runs": parse_logs()}
|
||||
|
||||
runs_count = len(data["runs"])
|
||||
print(f"Parsed {runs_count} runs.")
|
||||
|
||||
with open(OUTPUT_FILE, "w") as f:
|
||||
json.dump(data, f, indent=2)
|
||||
print(f"Written to {OUTPUT_FILE}")
|
||||
@@ -0,0 +1,95 @@
|
||||
{
|
||||
"runs": [
|
||||
{
|
||||
"model": "Qwen/Qwen3-14B-AWQ",
|
||||
"model_clean": "Qwen/Qwen3-14B-AWQ",
|
||||
"env": "TP1",
|
||||
"gpu_config": "single",
|
||||
"quant": "AWQ",
|
||||
"params_b": 14.0,
|
||||
"name_params_b": 14.0,
|
||||
"backend": "vLLM",
|
||||
"error": false,
|
||||
"test": "Throughput",
|
||||
"tps_mean": 112.69232830266365
|
||||
},
|
||||
{
|
||||
"model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
|
||||
"model_clean": "meta-llama/Meta-Llama-3.1-8B-Instruct",
|
||||
"env": "TP1",
|
||||
"gpu_config": "single",
|
||||
"quant": "BF16",
|
||||
"params_b": 8.0,
|
||||
"name_params_b": 8.0,
|
||||
"backend": "vLLM",
|
||||
"error": false,
|
||||
"test": "Throughput",
|
||||
"tps_mean": 278.99494393048457
|
||||
},
|
||||
{
|
||||
"model": "google/gemma-3-12b-it",
|
||||
"model_clean": "google/gemma-3-12b-it",
|
||||
"env": "TP1",
|
||||
"gpu_config": "single",
|
||||
"quant": "BF16",
|
||||
"params_b": 12.0,
|
||||
"name_params_b": 12.0,
|
||||
"backend": "vLLM",
|
||||
"error": false,
|
||||
"test": "Throughput",
|
||||
"tps_mean": 162.71078485804028
|
||||
},
|
||||
{
|
||||
"model": "dazipe/Qwen3-Next-80B-A3B-Instruct-GPTQ-Int4A16",
|
||||
"model_clean": "dazipe/Qwen3-Next-80B-A3B-Instruct-GPTQ-Int4A16",
|
||||
"env": "TP1",
|
||||
"gpu_config": "single",
|
||||
"quant": "GPTQ",
|
||||
"params_b": 80.0,
|
||||
"name_params_b": 80.0,
|
||||
"backend": "vLLM",
|
||||
"error": false,
|
||||
"test": "Throughput",
|
||||
"tps_mean": 112.62418795067208
|
||||
},
|
||||
{
|
||||
"model": "openai/gpt-oss-20b",
|
||||
"model_clean": "openai/gpt-oss-20b",
|
||||
"env": "TP1",
|
||||
"gpu_config": "single",
|
||||
"quant": "BF16",
|
||||
"params_b": 20.0,
|
||||
"name_params_b": 20.0,
|
||||
"backend": "vLLM",
|
||||
"error": false,
|
||||
"test": "Throughput",
|
||||
"tps_mean": 313.85817605876395
|
||||
},
|
||||
{
|
||||
"model": "cpatonn/Qwen3-Coder-30B-A3B-Instruct-GPTQ-4bit",
|
||||
"model_clean": "cpatonn/Qwen3-Coder-30B-A3B-Instruct-GPTQ-4bit",
|
||||
"env": "TP1",
|
||||
"gpu_config": "single",
|
||||
"quant": "GPTQ",
|
||||
"params_b": 30.0,
|
||||
"name_params_b": 30.0,
|
||||
"backend": "vLLM",
|
||||
"error": false,
|
||||
"test": "Throughput",
|
||||
"tps_mean": 271.7264154071495
|
||||
},
|
||||
{
|
||||
"model": "openai/gpt-oss-120b",
|
||||
"model_clean": "openai/gpt-oss-120b",
|
||||
"env": "TP1",
|
||||
"gpu_config": "single",
|
||||
"quant": "BF16",
|
||||
"params_b": 120.0,
|
||||
"name_params_b": 120.0,
|
||||
"backend": "vLLM",
|
||||
"error": false,
|
||||
"test": "Throughput",
|
||||
"tps_mean": 109.73523843987172
|
||||
}
|
||||
]
|
||||
}
|
||||
@@ -89,11 +89,11 @@ echo
|
||||
printf 'Machine: %s\n' "$MACHINE"
|
||||
printf 'GPU : %s\n\n' "$GPU"
|
||||
printf 'Repo : https://github.com/kyuz0/amd-strix-halo-vllm-toolboxes\n'
|
||||
printf 'Image : docker.io/kyuz0/vllm-therock-gfx1151-aotriton:latest\n\n'
|
||||
printf 'Image : docker.io/kyuz0/vllm-therock-gfx1151:latest\n\n'
|
||||
printf 'Included:\n'
|
||||
printf ' - %-16s → %s\n' "start-vllm (wizard)" "Beginner-friendly launcher that guides you through model & settings"
|
||||
printf ' - %-16s → %s\n' "vLLM server" "vllm serve Qwen/Qwen2.5-7B-Instruct --download-dir ~/vllm-models"
|
||||
printf ' - %-16s → %s\n' "API test" "curl localhost:8000/v1/chat/completions (see README)"
|
||||
printf ' - %-16s → %s\n' "start-vllm (TUI)" "Interactive launcher: Model select, Multi-GPU & Cache handling"
|
||||
printf ' - %-16s → %s\n' "vLLM server" "vllm serve meta-llama/Meta-Llama-3.1-8B-Instruct"
|
||||
printf ' - %-16s → %s\n' "API test" "curl localhost:8000/v1/chat/completions"
|
||||
echo
|
||||
printf 'SSH tip: ssh -L 8000:localhost:8000 user@host\n\n'
|
||||
|
||||
|
||||
@@ -1,54 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
set -euo pipefail
|
||||
|
||||
# Defaults (override via env: HOST, PORT, DOWNLOAD_DIR, EXTRA_FLAGS)
|
||||
HOST="${HOST:-0.0.0.0}"
|
||||
PORT="${PORT:-8000}"
|
||||
DOWNLOAD_DIR="${DOWNLOAD_DIR:-$HOME/vllm-models}"
|
||||
EXTRA_FLAGS="${EXTRA_FLAGS:-}"
|
||||
|
||||
models=(
|
||||
"Llama 2 7B Chat|meta-llama/Llama-2-7b-chat-hf|"
|
||||
"Qwen2.5 7B Instruct|Qwen/Qwen2.5-7B-Instruct|"
|
||||
"Qwen3 30B A3B Instruct|Qwen/Qwen3-30B-A3B-Instruct-2507|"
|
||||
"Qwen3 14B AWQ|Qwen/Qwen3-14B-AWQ|--quantization awq --dtype float16 --enforce-eager"
|
||||
"Gemma 3 27B instruct|google/gemma-3-27b-it|"
|
||||
"Gemma 3 12B Instruct|google/gemma-3-12b-it|"
|
||||
"Gemma 3 4B Instruct|google/gemma-3-4b-it|"
|
||||
)
|
||||
|
||||
echo "Select a model:"
|
||||
for i in "${!models[@]}"; do
|
||||
name="${models[$i]%%|*}"
|
||||
printf " [%d] %s\n" "$((i+1))" "$name"
|
||||
done
|
||||
|
||||
read -rp "Enter number: " choice
|
||||
[[ "$choice" =~ ^[1-9][0-9]*$ ]] || { echo "Invalid choice."; exit 1; }
|
||||
idx=$((choice-1))
|
||||
(( idx >= 0 && idx < ${#models[@]} )) || { echo "Invalid choice."; exit 1; }
|
||||
|
||||
IFS='|' read -r label repo flags <<< "${models[$idx]}"
|
||||
|
||||
mkdir -p "$DOWNLOAD_DIR"
|
||||
|
||||
CMD=(vllm serve "$repo" --host "$HOST" --port "$PORT" --download-dir "$DOWNLOAD_DIR")
|
||||
|
||||
# Per-model flags
|
||||
if [[ -n "${flags:-}" ]]; then
|
||||
# shellcheck disable=SC2206
|
||||
CMD+=($flags)
|
||||
fi
|
||||
|
||||
# Optional global extras: e.g. EXTRA_FLAGS="--gpu-memory-utilization 0.8"
|
||||
if [[ -n "${EXTRA_FLAGS:-}" ]]; then
|
||||
# shellcheck disable=SC2206
|
||||
CMD+=($EXTRA_FLAGS)
|
||||
fi
|
||||
|
||||
echo -e "Running:\n\n ${CMD[@]}\n"
|
||||
echo "API test → curl -s http://localhost:${PORT}/v1/models | jq -r '.data[0].id'"
|
||||
echo "SSH tip → ssh -L ${PORT}:localhost:${PORT} user@host"
|
||||
echo
|
||||
|
||||
exec "${CMD[@]}"
|
||||
@@ -0,0 +1,315 @@
|
||||
#!/usr/bin/env python3
|
||||
import sys
|
||||
import os
|
||||
import json
|
||||
import shutil
|
||||
import tempfile
|
||||
import subprocess
|
||||
from pathlib import Path
|
||||
|
||||
# Add benchmarks dir to path to import config
|
||||
SCRIPT_DIR = Path(__file__).parent.resolve()
|
||||
BENCH_DIR = SCRIPT_DIR.parent / "benchmarks"
|
||||
OPT_DIR = Path("/opt")
|
||||
|
||||
# Check /opt first (Container), then local fallback
|
||||
if (OPT_DIR / "run_vllm_bench.py").exists():
|
||||
sys.path.append(str(OPT_DIR))
|
||||
else:
|
||||
sys.path.append(str(BENCH_DIR))
|
||||
|
||||
try:
|
||||
from run_vllm_bench import MODEL_TABLE, MODELS_TO_RUN
|
||||
except ImportError:
|
||||
print("Error: Could not import run_vllm_bench.py config.")
|
||||
sys.exit(1)
|
||||
|
||||
if (OPT_DIR / "max_context_results.json").exists():
|
||||
RESULTS_FILE = OPT_DIR / "max_context_results.json"
|
||||
else:
|
||||
RESULTS_FILE = BENCH_DIR / "max_context_results.json"
|
||||
HOST = os.getenv("HOST", "0.0.0.0")
|
||||
PORT = os.getenv("PORT", "8000")
|
||||
|
||||
def check_dependencies():
|
||||
if not shutil.which("dialog"):
|
||||
print("Error: 'dialog' is required. Please install it (apt-get install dialog).")
|
||||
sys.exit(1)
|
||||
|
||||
def detect_gpus():
|
||||
"""Detects AMD GPUs via rocm-smi or /dev/dri."""
|
||||
try:
|
||||
# Try rocm-smi first
|
||||
res = subprocess.run(["rocm-smi", "--showid", "--csv"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
|
||||
if res.returncode == 0:
|
||||
count = res.stdout.count("GPU")
|
||||
if count > 0: return count
|
||||
except: pass
|
||||
|
||||
# Fallback to /dev/dri/render*
|
||||
try:
|
||||
return len(list(Path("/dev/dri").glob("renderD*")))
|
||||
except:
|
||||
return 1
|
||||
|
||||
def get_verified_config(model_id, tp_size, max_seqs):
|
||||
"""
|
||||
Reads max_context_results.json to find the best verified configuration.
|
||||
Returns dict: {'ctx': int, 'util': float}
|
||||
"""
|
||||
default_config = {
|
||||
"ctx": int(MODEL_TABLE.get(model_id, {}).get("ctx", 8192)),
|
||||
"util": 0.90 # Safe default
|
||||
}
|
||||
|
||||
if not RESULTS_FILE.exists():
|
||||
return default_config
|
||||
|
||||
try:
|
||||
with open(RESULTS_FILE, "r") as f:
|
||||
data = json.load(f)
|
||||
|
||||
# Filter for Model + TP + Sequences
|
||||
matches = [r for r in data
|
||||
if r["model"] == model_id
|
||||
and r["tp"] == tp_size
|
||||
and r["max_seqs"] == max_seqs
|
||||
and r["status"] == "success"]
|
||||
|
||||
if not matches:
|
||||
# Fallback 1: Try finding match with SAME TP but ANY Sequences (e.g. 1) to get base context?
|
||||
# Actually, safer to fallback to default or try finding nearest sequence?
|
||||
# Let's try finding exact match first. If fail, return default.
|
||||
return default_config
|
||||
|
||||
# Sort by Util desc, then Context desc
|
||||
# We prefer higher utilization if available (performance), as long as it is verified success
|
||||
matches.sort(key=lambda x: (float(x["util"]), x["max_context_1_user"]), reverse=True)
|
||||
|
||||
best = matches[0]
|
||||
return {
|
||||
"ctx": best["max_context_1_user"],
|
||||
"util": float(best["util"])
|
||||
}
|
||||
|
||||
except Exception as e:
|
||||
return default_config
|
||||
|
||||
def run_dialog(args):
|
||||
"""Runs dialog and returns stderr (selection)."""
|
||||
with tempfile.NamedTemporaryFile(mode="w+") as tf:
|
||||
cmd = ["dialog"] + args
|
||||
try:
|
||||
subprocess.run(cmd, stderr=tf, check=True)
|
||||
tf.seek(0)
|
||||
return tf.read().strip()
|
||||
except subprocess.CalledProcessError:
|
||||
return None # User cancelled
|
||||
|
||||
def nuke_vllm_cache():
|
||||
"""Removes vLLM cache directory to fix potential graph/incompatibility issues."""
|
||||
cache = Path.home() / ".cache" / "vllm"
|
||||
if cache.exists():
|
||||
try:
|
||||
print(f"Clearing vLLM cache at {cache}...", end="", flush=True)
|
||||
subprocess.run(["rm", "-rf", str(cache)], check=True)
|
||||
cache.mkdir(parents=True, exist_ok=True)
|
||||
print(" Done.")
|
||||
time.sleep(1)
|
||||
except Exception as e:
|
||||
print(f" Failed: {e}")
|
||||
|
||||
def configure_and_launch(model_idx, gpu_count):
|
||||
model_id = MODELS_TO_RUN[model_idx]
|
||||
config = MODEL_TABLE[model_id]
|
||||
|
||||
# Static Config
|
||||
valid_tps = config.get("valid_tp", [1])
|
||||
max_tp = max(valid_tps) if valid_tps else 1
|
||||
|
||||
# Defaults
|
||||
current_tp = min(gpu_count, max_tp)
|
||||
current_seqs = 1 # Default to 1 concurrent user/request for stability
|
||||
|
||||
# Initial Lookup
|
||||
verified = get_verified_config(model_id, current_tp, current_seqs)
|
||||
current_ctx = verified["ctx"]
|
||||
current_util = verified["util"]
|
||||
|
||||
clear_cache = False
|
||||
use_eager = config.get("enforce_eager", False) # Default to model config, usually False
|
||||
use_rocm_attn = False # Default to Triton
|
||||
|
||||
name = model_id.split("/")[-1]
|
||||
|
||||
while True:
|
||||
cache_status = "YES" if clear_cache else "NO"
|
||||
eager_status = "YES" if use_eager else "NO"
|
||||
attn_backend = "ROCm" if use_rocm_attn else "Triton"
|
||||
|
||||
menu_args = [
|
||||
"--clear", "--backtitle", f"AMD R9700 vLLM Launcher (GPUs: {gpu_count})",
|
||||
"--title", f"Configuration: {name}",
|
||||
"--menu", "Customize Launch Parameters:", "22", "65", "9",
|
||||
"1", f"Tensor Parallelism: {current_tp}",
|
||||
"2", f"Concurrent Requests: {current_seqs}",
|
||||
"3", f"Context Length: {current_ctx} (Verified)",
|
||||
"4", f"GPU Utilization: {current_util} (Verified)",
|
||||
"5", f"Attention Backend: {attn_backend}",
|
||||
"6", f"Erase vLLM Cache: {cache_status}",
|
||||
"7", f"Force Eager Mode: {eager_status}",
|
||||
"8", "LAUNCH SERVER"
|
||||
]
|
||||
|
||||
choice = run_dialog(menu_args)
|
||||
if not choice: return False # Back/Cancel
|
||||
|
||||
if choice == "1":
|
||||
# TP Selection
|
||||
new_tp = run_dialog([
|
||||
"--title", "Tensor Parallelism",
|
||||
"--rangebox", f"Set TP Size (1-{max_tp})", "10", "40", "1", str(max_tp), str(current_tp)
|
||||
])
|
||||
if new_tp:
|
||||
new_tp_int = int(new_tp)
|
||||
if new_tp_int != current_tp:
|
||||
current_tp = new_tp_int
|
||||
# RE-CALCULATE Config
|
||||
verified = get_verified_config(model_id, current_tp, current_seqs)
|
||||
current_ctx = verified["ctx"]
|
||||
current_util = verified["util"]
|
||||
|
||||
elif choice == "2":
|
||||
# Max Seqs Selection
|
||||
new_seqs = run_dialog([
|
||||
"--title", "Concurrent Requests",
|
||||
"--menu", "Select Max Concurrent Requests:", "12", "40", "4",
|
||||
"1", "1 (Latency Focus)",
|
||||
"4", "4 (Balanced)",
|
||||
"8", "8 (Throughput)",
|
||||
"16", "16 (Max Load)"
|
||||
])
|
||||
if new_seqs:
|
||||
current_seqs = int(new_seqs)
|
||||
# RE-CALCULATE Config based on new concurrency
|
||||
verified = get_verified_config(model_id, current_tp, current_seqs)
|
||||
current_ctx = verified["ctx"]
|
||||
current_util = verified["util"]
|
||||
|
||||
elif choice == "3":
|
||||
# Configured Length Override
|
||||
new_ctx = run_dialog([
|
||||
"--title", "Context Length",
|
||||
"--inputbox", f"Override verified limit ({current_ctx}):", "10", "40", str(current_ctx)
|
||||
])
|
||||
if new_ctx: current_ctx = int(new_ctx)
|
||||
|
||||
elif choice == "4":
|
||||
# Util Override
|
||||
pass
|
||||
|
||||
elif choice == "5":
|
||||
# Toggle Attention Backend
|
||||
use_rocm_attn = not use_rocm_attn
|
||||
|
||||
elif choice == "6":
|
||||
# Toggle Cache
|
||||
if not clear_cache:
|
||||
# Enabling it -> Show Warning
|
||||
warn_msg = (
|
||||
"WARNING: Erasing the vLLM cache will remove the compiled compute graphs.\n\n"
|
||||
"This is useful if you are experiencing crashes, 'invalid graph' errors,\n"
|
||||
"or have switched vLLM versions recently.\n\n"
|
||||
"However, the next startup will take longer as graphs are re-compiled.\n\n"
|
||||
"Are you sure you want to enable this?"
|
||||
)
|
||||
confirm = run_dialog([
|
||||
"--title", "Erase Cache Warning",
|
||||
"--yesno", warn_msg, "12", "60"
|
||||
])
|
||||
|
||||
# If confirm is not None (exit 0), it is YES.
|
||||
if confirm is not None:
|
||||
clear_cache = True
|
||||
else:
|
||||
# Disabling it -> No warning needed
|
||||
clear_cache = False
|
||||
|
||||
elif choice == "7":
|
||||
# Toggle Eager Mode
|
||||
use_eager = not use_eager
|
||||
|
||||
elif choice == "8":
|
||||
# Launch
|
||||
break
|
||||
|
||||
# Build Command
|
||||
subprocess.run(["clear"])
|
||||
|
||||
if clear_cache:
|
||||
nuke_vllm_cache()
|
||||
|
||||
cmd = [
|
||||
"vllm", "serve", model_id,
|
||||
"--host", HOST,
|
||||
"--port", PORT,
|
||||
"--tensor-parallel-size", str(current_tp),
|
||||
"--max-num-seqs", str(current_seqs),
|
||||
"--max-model-len", str(current_ctx),
|
||||
"--gpu-memory-utilization", str(current_util),
|
||||
"--dtype", "auto"
|
||||
]
|
||||
|
||||
if config.get("trust_remote"): cmd.append("--trust-remote-code")
|
||||
if use_eager: cmd.append("--enforce-eager")
|
||||
|
||||
# Env Vars
|
||||
env = os.environ.copy()
|
||||
env.update(config.get("env", {}))
|
||||
|
||||
if use_rocm_attn:
|
||||
env["VLLM_V1_USE_PREFILL_DECODE_ATTENTION"] = "1"
|
||||
env["VLLM_USE_TRITON_FLASH_ATTN"] = "0"
|
||||
# Optional: Explicitly mention these in print
|
||||
|
||||
|
||||
print("\n" + "="*60)
|
||||
print(f" Launching: {name}")
|
||||
print(f" Config: TP={current_tp} | Seqs={current_seqs} | Ctx={current_ctx} | Util={current_util}")
|
||||
print(f" Backend: {'ROCm' if use_rocm_attn else 'Triton'}")
|
||||
if clear_cache:
|
||||
print(f" Action: Clearing vLLM Cache (~/.cache/vllm)")
|
||||
print(f" Command: {' '.join(cmd)}")
|
||||
print("="*60 + "\n")
|
||||
|
||||
os.execvpe("vllm", cmd, env)
|
||||
|
||||
def main():
|
||||
check_dependencies()
|
||||
gpu_count = detect_gpus()
|
||||
|
||||
while True:
|
||||
# Build Model Menu
|
||||
menu_items = []
|
||||
for i, m_id in enumerate(MODELS_TO_RUN):
|
||||
name = m_id.split("/")[-1]
|
||||
# Pre-calc verified ctx for 'default' TP to show in menu?
|
||||
# Or just show names. Just names is cleaner.
|
||||
config = MODEL_TABLE[m_id]
|
||||
menu_items.extend([str(i), name])
|
||||
|
||||
choice = run_dialog([
|
||||
"--clear", "--backtitle", f"AMD R9700 vLLM Launcher (GPUs: {gpu_count})",
|
||||
"--title", "Select Model",
|
||||
"--menu", "Choose a model to serve:", "20", "60", "10"
|
||||
] + menu_items)
|
||||
|
||||
if not choice:
|
||||
subprocess.run(["clear"])
|
||||
print("Selection cancelled.")
|
||||
sys.exit(0)
|
||||
|
||||
configure_and_launch(int(choice), gpu_count)
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
新增問題並參考
封鎖使用者