remove cmd_stop configuration and functionality from PR #40 (#44 )

* remove cmd_stop functionality from #40
Use client request context in proxy request (#43 )
2025-01-31 12:42:44 -08:00 · 2025-01-31 10:21:49 -08:00 · 2025-01-31 10:09:07 -08:00 · 2025-01-30 16:59:57 -08:00 · 2025-01-20 14:39:52 -08:00 · 2025-01-16 12:06:38 -08:00
29 changed files with 1419 additions and 172 deletions
@@ -30,4 +30,4 @@ jobs:
          version: '~> v2'
          args: release --clean
        env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
@@ -2,4 +2,5 @@
 .env
 build/
 dist/
-.vscode
+.vscode
+.DS_Store
@@ -2,6 +2,16 @@
 APP_NAME = llama-swap
 BUILD_DIR = build

+# Get the current Git hash
+GIT_HASH := $(shell git rev-parse --short HEAD)
+ifneq ($(shell git status --porcelain),)
+    # There are untracked changes
+    GIT_HASH := $(GIT_HASH)+
+endif
+
+# Capture the current build date in RFC3339 format
+BUILD_DATE := $(shell date -u +"%Y-%m-%dT%H:%M:%SZ")
+
 # Default target: Builds binaries for both OSX and Linux
 all: mac linux simple-responder

@@ -18,12 +28,12 @@ test-all:
 # Build OSX binary
 mac:
 	@echo "Building Mac binary..."
-	GOOS=darwin GOARCH=arm64 go build -o $(BUILD_DIR)/$(APP_NAME)-darwin-arm64
+	GOOS=darwin GOARCH=arm64 go build -ldflags="-X main.commit=${GIT_HASH} -X main.version=local_${GIT_HASH} -X main.date=${BUILD_DATE}" -o $(BUILD_DIR)/$(APP_NAME)-darwin-arm64

 # Build Linux binary
 linux:
 	@echo "Building Linux binary..."
-	GOOS=linux GOARCH=amd64 go build -o $(BUILD_DIR)/$(APP_NAME)-linux-amd64
+	GOOS=linux GOARCH=amd64 go build -ldflags="-X main.commit=${GIT_HASH} -X main.version=local_${GIT_HASH} -X main.date=${BUILD_DATE}" -o $(BUILD_DIR)/$(APP_NAME)-linux-amd64

 # for testing proxy.Process
 simple-responder:
@@ -35,5 +45,19 @@ simple-responder:
 $(BUILD_DIR):
 	mkdir -p $(BUILD_DIR)

+# Create a new release tag
+release:
+	@echo "Checking for unstaged changes..."
+	@if [ -n "$(shell git status --porcelain)" ]; then \
+		echo "Error: There are unstaged changes. Please commit or stash your changes before creating a release tag." >&2; \
+		exit 1; \
+	fi
+
+# Get the highest tag in v{number} format, increment it, and create a new tag
+	@highest_tag=$$(git tag --sort=-v:refname | grep -E '^v[0-9]+$$' | head -n 1 || echo "v0"); \
+	new_tag="v$$(( $${highest_tag#v} + 1 ))"; \
+	echo "tagging new version: $$new_tag"; \
+	git tag "$$new_tag";
+
 # Phony targets
 .PHONY: all clean osx linux
@@ -2,18 +2,40 @@

 ![llama-swap header image](header.jpeg)

-llama-swap is a golang server that automatically swaps the llama.cpp server on demand. Since [llama.cpp's server](https://github.com/ggerganov/llama.cpp/tree/master/examples/server) can't swap models, let's swap the server instead!
+# Introduction
+llama-swap is a light weight, transparent proxy server that provides automatic model swapping to llama.cpp's server.

-Features:
+Written in golang, it is very easy to install (single binary with no dependancies) and configure (single yaml file).
+
+Download a pre-built [release](https://github.com/mostlygeek/llama-swap/releases) or build it yourself from source with `make clean all`.
+
+## How does it work?
+When a request is made to an OpenAI compatible endpoint, lama-swap will extract the `model` value and load the appropriate server configuration to serve it. If a server is already running it will stop it and start the correct one. This is where the "swap" part comes in. The upstream server is automatically swapped to the correct one to serve the request.
+
+In the most basic configuration llama-swap handles one model at a time. For more advanced use cases, the `profiles` feature can load multiple models at the same time. You have complete control over how your system resources are used.
+
+## Do I need to use llama.cpp's server (llama-server)?
+Any OpenAI compatible server would work. llama-swap was originally designed for llama-server and it is the best supported. For Python based inference servers like vllm or tabbyAPI it is recommended to run them via podman. This provides clean environment isolation as well as responding correctly to `SIGTERM` signals to shutdown.
+
+## Features:

 - ✅ Easy to deploy: single binary with no dependencies
- ✅ Single yaml configuration file
- ✅ Automatically switching between models
- ✅ Full control over llama.cpp server settings per model
- ✅ OpenAI API support (`v1/completions` and `v1/chat/completions`)
+- ✅ Easy to config: single yaml file
+- ✅ On-demand model switching
+- ✅ Full control over server settings per model
+- ✅ OpenAI API supported endpoints:
+  - `v1/completions`
+  - `v1/chat/completions`
+  - `v1/embeddings`
+  - `v1/rerank`
+  - `v1/audio/speech` ([#36](https://github.com/mostlygeek/llama-swap/issues/36))
 - ✅ Multiple GPU support
+- ✅ Docker and Podman support
 - ✅ Run multiple models at once with `profiles`
 - ✅ Remote log monitoring at `/log`
+- ✅ Automatic unloading of models from GPUs after timeout
+- ✅ Use any local OpenAI compatible server (llama.cpp, vllm, tabbyAPI, etc)
+- ✅ Direct access to upstream HTTP server via `/upstream/:model_id` ([demo](https://github.com/mostlygeek/llama-swap/pull/31))

 ## config.yaml

@@ -24,6 +46,9 @@ llama-swap's configuration is purposefully simple.
 # Default (and minimum) is 15 seconds
 healthCheckTimeout: 60

+# Write HTTP logs (useful for troubleshooting), defaults to false
+logRequests: true
+
 # define valid model values and the upstream server start
 models:
  "llama":
@@ -59,11 +84,26 @@ models:
      --model path/to/Qwen2.5-1.5B-Instruct-Q4_K_M.gguf
    proxy: http://127.0.0.1:8999

+  # unlisted models do not show up in /v1/models or /upstream lists
+  # but they can still be requested as normal
+  "qwen-unlisted":
+    cmd: llama-server --port 9999 -m Llama-3.2-1B-Instruct-Q4_K_M.gguf -ngl 0
+    unlisted: true
+
+  # Docker Support (v26.1.4+ required!)
+  "docker-llama":
+    proxy: "http://127.0.0.1:9790"
+    cmd: >
+      docker run --name dockertest
+      --init --rm -p 9790:8080 -v /mnt/nvme/models:/models
+      ghcr.io/ggerganov/llama.cpp:server
+      --model '/models/Qwen2.5-Coder-0.5B-Instruct-Q4_K_M.gguf'
+
 # profiles make it easy to managing multi model (and gpu) configurations.
 #
 # Tips:
 #  - each model must be listening on a unique address and port
-#  - the model name is in this format: "profile_name/model", like "coding/qwen"
+#  - the model name is in this format: "profile_name:model", like "coding:qwen"
 #  - the profile will load and unload all models in the profile at the same time
 profiles:
  coding:
@@ -71,31 +111,44 @@ profiles:
    - "llama"
 ```

-## Installation
+### Advanced Examples
+
+- [config.example.yaml](config.example.yaml) includes example for supporting `v1/embeddings` and `v1/rerank` endpoints
+- [Speculative Decoding](examples/speculative-decoding/README.md) - using a small draft model can increase inference speeds from 20% to 40%. This example includes a configurations Qwen2.5-Coder-32B (2.5x increase) and Llama-3.1-70B (1.4x increase) in the best cases.
+- [Optimizing Code Generation](examples/benchmark-snakegame/README.md) - find the optimal settings for your machine. This example demonstrates defining multiple configurations and testing which one is fastest.
+
+### Installation

 1. Create a configuration file, see [config.example.yaml](config.example.yaml)
 1. Download a [release](https://github.com/mostlygeek/llama-swap/releases) appropriate for your OS and architecture.
    * _Note: Windows currently untested._
 1. Run the binary with `llama-swap --config path/to/config.yaml`

+### Building from source
+
+1. Install golang for your system
+1. `git clone git@github.com:mostlygeek/llama-swap.git`
+1. `make clean all`
+1. Binaries will be in `build/` subdirectory
+
 ## Monitoring Logs

-The `/logs` endpoint is available to monitor what llama-swap is doing. It will send the last 10KB of logs. Useful for monitoring the output of llama-server. It also supports streaming of logs.
+Open the `http://<host>/logs` with your browser to get a web interface with streaming logs.

-Usage:
+Of course, CLI access is also supported:

 ```
 # sends up to the last 10KB of logs
 curl http://host/logs'

-# streams logs using chunk encoding
+# streams logs
 curl -Ns 'http://host/logs/stream'

+# stream and filter logs with linux pipes
+curl -Ns http://host/logs/stream | grep 'eval time'
+
 # skips history and just streams new log entries
 curl -Ns 'http://host/logs/stream?no-history'
-
-# streams logs using Server Sent Events
-curl -Ns 'http://host/logs/streamSSE'
 ```

 ## Systemd Unit Files
@@ -122,9 +175,3 @@ StartLimitInterval=30
 [Install]
 WantedBy=multi-user.target
 ```
-
-## Building from Source
-
-1. Install golang for your system
-1. run `make clean all`
-1. binaries will be built into `build/` directory
@@ -2,6 +2,9 @@
 # Default (and minimum): 15 seconds
 healthCheckTimeout: 15

+# Log HTTP requests helpful for troubleshoot, defaults to False
+logRequests: true
+
 models:
  "llama":
    cmd: >
@@ -26,6 +29,39 @@ models:
    aliases:
    - gpt-3.5-turbo

+  # Embedding example with Nomic
+  # https://huggingface.co/nomic-ai/nomic-embed-text-v1.5-GGUF
+  "nomic":
+    proxy: http://127.0.0.1:9005
+    cmd: >
+      models/llama-server-osx --port 9005
+      -m models/nomic-embed-text-v1.5.Q8_0.gguf
+      --ctx-size 8192
+      --batch-size 8192
+      --rope-scaling yarn
+      --rope-freq-scale 0.75
+      -ngl 99
+      --embeddings
+
+  # Reranking example with bge-reranker
+  # https://huggingface.co/gpustack/bge-reranker-v2-m3-GGUF
+  "bge-reranker":
+    proxy: http://127.0.0.1:9006
+    cmd: >
+      models/llama-server-osx --port 9006
+      -m models/bge-reranker-v2-m3-Q4_K_M.gguf
+      --ctx-size 8192
+      --reranking
+
+  # Docker Support (v26.1.4+ required!)
+  "dockertest":
+    proxy: "http://127.0.0.1:9790"
+    cmd: >
+      docker run --name dockertest
+      --init --rm -p 9790:8080 -v /mnt/nvme/models:/models
+      ghcr.io/ggerganov/llama.cpp:server
+      --model '/models/Qwen2.5-Coder-0.5B-Instruct-Q4_K_M.gguf'
+
  "simple":
    # example of setting environment variables
    env:
@@ -33,6 +69,7 @@ models:
      - env1=hello
    cmd: build/simple-responder --port 8999
    proxy: http://127.0.0.1:8999
+    unlisted: true

    # use "none" to skip check. Caution this may cause some requests to fail
    # until the upstream server is ready for traffic
@@ -42,9 +79,11 @@ models:
  "broken":
    cmd: models/llama-server-osx --port 8999 -m models/doesnotexist.gguf
    proxy: http://127.0.0.1:8999
+    unlisted: true
  "broken_timeout":
    cmd: models/llama-server-osx --port 8999 -m models/qwen2.5-0.5b-instruct-q8_0.gguf
    proxy: http://127.0.0.1:9000
+    unlisted: true

 # creating a coding profile with models for code generation and general questions
 profiles:
@@ -0,0 +1,6 @@
+# Example Configs and Use Cases
+
+A collections of usecases and examples for getting the most out of llama-swap.
+
+* [Speculative Decoding](speculative-decoding/README.md) - using a small draft model can increase inference speeds from 20% to 40%. This example includes a configurations Qwen2.5-Coder-32B (2.5x increase) and Llama-3.1-70B (1.4x increase) in the best cases.
+* [Optimizing Code Generation](benchmark-snakegame/README.md) - find the optimal settings for your machine. This example demonstrates defining multiple configurations and testing which one is fastest.
@@ -0,0 +1,123 @@
+# Optimizing Code Generation with llama-swap
+
+Finding the best mix of settings for your hardware can be time consuming. This example demonstrates using a custom configuration file to automate testing different scenarios to find the an optimal configuration.
+
+The benchmark writes a snake game in Python, TypeScript, and Swift using the Qwen 2.5 Coder models. The experiments were done using a 3090 and a P40.
+
+**Benchmark Scenarios**
+
+Three scenarios are tested:
+
+- 3090-only: Just the main model on the 3090
+- 3090-with-draft: the main and draft models on the 3090
+- 3090-P40-draft: the main model on the 3090 with the draft model offloaded to the P40
+
+**Available Devices**
+
+Use the following command to list available devices IDs for the configuration:
+
+```
+$ /mnt/nvme/llama-server/llama-server-f3252055 --list-devices
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 4 CUDA devices:
+  Device 0: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
+  Device 1: Tesla P40, compute capability 6.1, VMM: yes
+  Device 2: Tesla P40, compute capability 6.1, VMM: yes
+  Device 3: Tesla P40, compute capability 6.1, VMM: yes
+Available devices:
+  CUDA0: NVIDIA GeForce RTX 3090 (24154 MiB, 406 MiB free)
+  CUDA1: Tesla P40 (24438 MiB, 22942 MiB free)
+  CUDA2: Tesla P40 (24438 MiB, 24144 MiB free)
+  CUDA3: Tesla P40 (24438 MiB, 24144 MiB free)
+```
+
+**Configuration**
+
+The configuration file, `benchmark-config.yaml`, defines the three scenarios:
+
+```yaml
+models:
+  "3090-only":
+    proxy: "http://127.0.0.1:9503"
+    cmd: >
+      /mnt/nvme/llama-server/llama-server-f3252055
+      --host 127.0.0.1 --port 9503
+      --flash-attn
+      --slots
+
+      --model /mnt/nvme/models/Qwen2.5-Coder-32B-Instruct-Q4_K_M.gguf
+      -ngl 99
+      --device CUDA0
+
+      --ctx-size 32768
+      --cache-type-k q8_0 --cache-type-v q8_0
+
+  "3090-with-draft":
+    proxy: "http://127.0.0.1:9503"
+    # --ctx-size 28500 max that can fit on 3090 after draft model
+    cmd: >
+      /mnt/nvme/llama-server/llama-server-f3252055
+      --host 127.0.0.1 --port 9503
+      --flash-attn
+      --slots
+
+      --model /mnt/nvme/models/Qwen2.5-Coder-32B-Instruct-Q4_K_M.gguf
+      -ngl 99
+      --device CUDA0
+
+      --model-draft /mnt/nvme/models/Qwen2.5-Coder-0.5B-Instruct-Q8_0.gguf
+      -ngld 99
+      --draft-max 16
+      --draft-min 4
+      --draft-p-min 0.4
+      --device-draft CUDA0
+
+      --ctx-size 28500
+      --cache-type-k q8_0 --cache-type-v q8_0
+
+  "3090-P40-draft":
+    proxy: "http://127.0.0.1:9503"
+    cmd: >
+      /mnt/nvme/llama-server/llama-server-f3252055
+      --host 127.0.0.1 --port 9503
+      --flash-attn --metrics
+      --slots
+      --model /mnt/nvme/models/Qwen2.5-Coder-32B-Instruct-Q4_K_M.gguf
+      -ngl 99
+      --device CUDA0
+
+      --model-draft /mnt/nvme/models/Qwen2.5-Coder-0.5B-Instruct-Q8_0.gguf
+      -ngld 99
+      --draft-max 16
+      --draft-min 4
+      --draft-p-min 0.4
+      --device-draft CUDA1
+
+      --ctx-size 32768
+      --cache-type-k q8_0 --cache-type-v q8_0
+```
+
+> Note in the `3090-with-draft` scenario the `--ctx-size` had to be reduced from 32768 to to accommodate the draft model.
+
+
+**Running the Benchmark**
+
+To run the benchmark, execute the following commands:
+
+1. `llama-swap -config benchmark-config.yaml`
+1. `./run-benchmark.sh http://localhost:8080 "3090-only" "3090-with-draft" "3090-P40-draft"`
+
+The [benchmark script](run-benchmark.sh) generates a CSV output of the results, which can be converted to a Markdown table for readability.
+
+**Results (tokens/second)**
+
+| model           | python | typescript | swift |
+|-----------------|--------|------------|-------|
+| 3090-only       | 34.03  | 34.01      | 34.01 |
+| 3090-with-draft | 106.65 | 70.48      | 57.89 |
+| 3090-P40-draft  | 81.54  | 60.35      | 46.50 |
+
+Many different factors, like the programming language, can have big impacts on the performance gains. However, with a custom configuration file for benchmarking it is easy to test the different variations to discover what's best for your hardware.
+
+Happy coding!
@@ -0,0 +1,40 @@
+#!/usr/bin/env bash
+
+# This script generates a CSV file showing the token/second for generating a Snake Game in python, typescript and swift
+# It was created to test the effects of speculative decoding and the various draft settings on performance.
+#
+# Writing code with a low temperature seems to provide fairly consistent logic.
+#
+# Usage: ./benchmark.sh <url> <model1> [model2 ...]
+# Example: ./benchmark.sh http://localhost:8080 model1 model2
+
+if [ "$#" -lt 2 ]; then
+    echo "Usage: $0 <url> <model1> [model2 ...]"
+    exit 1
+fi
+
+url=$1; shift
+
+echo "model,python,typescript,swift"
+
+for model in "$@"; do
+
+    echo -n "$model,"
+
+    for lang in "python" "typescript" "swift"; do
+        # expects a llama.cpp after PR https://github.com/ggerganov/llama.cpp/pull/10548
+        # (Dec 3rd/2024)
+        time=$(curl -s --url "$url/v1/chat/completions" -d "{\"messages\": [{\"role\": \"system\", \"content\": \"you only write code.\"}, {\"role\": \"user\", \"content\": \"write snake game in $lang\"}], \"top_k\": 1, \"timings_per_token\":true, \"model\":\"$model\"}" | jq -r .timings.predicted_per_second)
+
+        if [ $? -ne 0 ]; then
+            time="error"
+            exit 1
+        fi
+
+        if [ "$lang" != "swift" ]; then
+            printf "%0.2f tps," $time
+        else
+            printf "%0.2f tps\n" $time
+        fi
+    done
+done
@@ -0,0 +1,124 @@
+# Speculative Decoding
+
+Speculative decoding can significantly improve the tokens per second. However, this comes at the cost of increased VRAM usage for the draft model. The examples provided are based on a server with three P40s and one 3090.
+
+## Coding Use Case
+
+This example uses Qwen2.5 Coder 32B with the 0.5B model as a draft. A quantization of Q8_0 was chosen for the draft model, as quantization has a greater impact on smaller models.
+
+The models used are:
+
+* [Bartowski Qwen2.5-Coder-32B-Instruct](https://huggingface.co/bartowski/Qwen2.5-Coder-32B-Instruct-GGUF)
+* [Bartowski Qwen2.5-Coder-0.5B-Instruct](https://huggingface.co/bartowski/Qwen2.5-Coder-0.5B-Instruct-GGUF)
+
+The llama-swap configuration is as follows:
+
+```yaml
+models:
+  "qwen-coder-32b-q4":
+    # main model on 3090, draft on P40 #1
+    cmd: >
+      /mnt/nvme/llama-server/llama-server-be0e35
+      --host 127.0.0.1 --port 9503
+      --flash-attn --metrics
+      --slots
+      --model /mnt/nvme/models/Qwen2.5-Coder-32B-Instruct-Q4_K_M.gguf
+      -ngl 99
+      --ctx-size 19000
+      --model-draft /mnt/nvme/models/Qwen2.5-Coder-0.5B-Instruct-Q8_0.gguf
+      -ngld 99
+      --draft-max 16
+      --draft-min 4
+      --draft-p-min 0.4
+      --device CUDA0
+      --device-draft CUDA1
+    proxy: "http://127.0.0.1:9503"
+```
+
+In this configuration, two GPUs are used: a 3090 (CUDA0) for the main model and a P40 (CUDA1) for the draft model. Although both models can fit on the 3090, relocating the draft model to the P40 freed up space for a larger context size. Despite the P40 being about 1/3rd the speed of the 3090, the small model still improved tokens per second.
+
+Multiple tests were run with various parameters, and the fastest result was chosen for the configuration. In all tests, the 0.5B model produced the largest improvements to tokens per second.
+
+Baseline: 33.92 tokens/second on 3090 without a draft model.
+
+| draft-max | draft-min | draft-p-min | python | TS | swift |
+|-----------|-----------|-------------|--------|----|-------|
+| 16 | 1 | 0.9 | 71.64 | 55.55 | 48.06 |
+| 16 | 1 | 0.4 | 83.21 | 58.55 | 45.50 |
+| 16 | 1 | 0.1 | 79.72 | 55.66 | 43.94 |
+| 16 | 2 | 0.9 | 68.47 | 55.13 | 43.12 |
+| 16 | 2 | 0.4 | 82.82 | 57.42 | 48.83 |
+| 16 | 2 | 0.1 | 81.68 | 51.37 | 45.72 |
+| 16 | 4 | 0.9 | 66.44 | 48.49 | 42.40 |
+| 16 | 4 | 0.4 | _83.62_ (fastest)| _58.29_ | _50.17_ |
+| 16 | 4 | 0.1 | 82.46 | 51.45 | 40.71 |
+| 8 | 1 | 0.4 | 67.07 | 55.17 | 48.46 |
+| 4 | 1 | 0.4 | 50.13 | 44.96 | 40.79 |
+
+The test script can be found in this [gist](https://gist.github.com/mostlygeek/da429769796ac8a111142e75660820f1). It is a simple curl script that prompts generating a snake game in Python, TypeScript, or Swift. Evaluation metrics were pulled from llama.cpp's logs.
+
+```bash
+for lang in "python" "typescript" "swift"; do
+    echo "Generating Snake Game in $lang using $model"
+    curl -s --url http://localhost:8080/v1/chat/completions -d "{\"messages\": [{\"role\": \"system\", \"content\": \"you only write code.\"}, {\"role\": \"user\", \"content\": \"write snake game in $lang\"}], \"temperature\": 0.1, \"model\":\"$model\"}" > /dev/null
+done
+```
+
+Python consistently outperformed Swift in all tests, likely due to the 0.5B draft model being more proficient in generating Python code accepted by the larger 32B model.
+
+## Chat
+
+This configuration is for a regular chat use case. It produces approximately 13 tokens/second in typical use, up from ~9 tokens/second with only the 3xP40s. This is great news for P40 owners.
+
+The models used are:
+
+* [Bartowski Meta-Llama-3.1-70B-Instruct-GGUF](https://huggingface.co/bartowski/Meta-Llama-3.1-70B-Instruct-GGUF)
+* [Bartowski Llama-3.2-3B-Instruct-GGUF](https://huggingface.co/bartowski/Llama-3.2-3B-Instruct-GGUF)
+
+```yaml
+models:
+  "llama-70B":
+    cmd: >
+      /mnt/nvme/llama-server/llama-server-be0e35
+      --host 127.0.0.1 --port 9602
+      --flash-attn --metrics
+      --split-mode row
+      --ctx-size 80000
+      --model /mnt/nvme/models/Meta-Llama-3.1-70B-Instruct-Q4_K_L.gguf
+      -ngl 99
+      --model-draft /mnt/nvme/models/Llama-3.2-3B-Instruct-Q4_K_M.gguf
+      -ngld 99
+      --draft-max 16
+      --draft-min 1
+      --draft-p-min 0.4
+      --device-draft CUDA0
+      --tensor-split 0,1,1,1
+```
+
+In this configuration, Llama-3.1-70B is split across three P40s, and Llama-3.2-3B is on the 3090.
+
+Some flags deserve further explanation:
+
+* `--split-mode row` - increases inference speeds using multiple P40s by about 30%. This is a P40-specific feature.
+* `--tensor-split 0,1,1,1` - controls how the main model is split across the GPUs. This means 0% on the 3090 and an even split across the P40s. A value of `--tensor-split 0,5,4,1` would mean 0% on the 3090, 50%, 40%, and 10% respectively across the other P40s. However, this would exceed the available VRAM.
+* `--ctx-size 80000` - the maximum context size that can fit in the remaining VRAM.
+
+## What is CUDA0, CUDA1, CUDA2, CUDA3?
+
+These devices are the IDs used by llama.cpp.
+
+```bash
+$ ./llama-server --list-devices
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 4 CUDA devices:
+  Device 0: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
+  Device 1: Tesla P40, compute capability 6.1, VMM: yes
+  Device 2: Tesla P40, compute capability 6.1, VMM: yes
+  Device 3: Tesla P40, compute capability 6.1, VMM: yes
+Available devices:
+  CUDA0: NVIDIA GeForce RTX 3090 (24154 MiB, 23892 MiB free)
+  CUDA1: Tesla P40 (24438 MiB, 24290 MiB free)
+  CUDA2: Tesla P40 (24438 MiB, 24290 MiB free)
+  CUDA3: Tesla P40 (24438 MiB, 24290 MiB free)
+```
@@ -20,6 +20,7 @@ require (
 	github.com/go-playground/universal-translator v0.18.1 // indirect
 	github.com/go-playground/validator/v10 v10.20.0 // indirect
 	github.com/goccy/go-json v0.10.2 // indirect
+	github.com/google/shlex v0.0.0-20191202100458-e7afc7fbc510 // indirect
 	github.com/json-iterator/go v1.1.12 // indirect
 	github.com/klauspost/cpuid/v2 v2.2.7 // indirect
 	github.com/leodido/go-urn v1.4.0 // indirect
@@ -31,9 +32,9 @@ require (
 	github.com/twitchyliquid64/golang-asm v0.15.1 // indirect
 	github.com/ugorji/go/codec v1.2.12 // indirect
 	golang.org/x/arch v0.8.0 // indirect
-	golang.org/x/crypto v0.23.0 // indirect
-	golang.org/x/net v0.25.0 // indirect
-	golang.org/x/sys v0.20.0 // indirect
-	golang.org/x/text v0.15.0 // indirect
+	golang.org/x/crypto v0.31.0 // indirect
+	golang.org/x/net v0.33.0 // indirect
+	golang.org/x/sys v0.28.0 // indirect
+	golang.org/x/text v0.21.0 // indirect
 	google.golang.org/protobuf v1.34.1 // indirect
 )
@@ -24,6 +24,8 @@ github.com/go-playground/validator/v10 v10.20.0/go.mod h1:dbuPbCMFw/DrkbEynArYaC
 github.com/goccy/go-json v0.10.2 h1:CrxCmQqYDkv1z7lO7Wbh2HN93uovUHgrECaO5ZrCXAU=
 github.com/goccy/go-json v0.10.2/go.mod h1:6MelG93GURQebXPDq3khkgXZkazVtN9CRI+MGFi0w8I=
 github.com/google/gofuzz v1.0.0/go.mod h1:dBl0BpW6vV/+mYPU4Po3pmUjxk6FQPldtuIdl/M65Eg=
+github.com/google/shlex v0.0.0-20191202100458-e7afc7fbc510 h1:El6M4kTTCOh6aBiKaUGG7oYTSPP8MxqL4YI3kZKwcP4=
+github.com/google/shlex v0.0.0-20191202100458-e7afc7fbc510/go.mod h1:pupxD2MaaD3pAXIBCelhxNneeOaAeabZDe5s4K6zSpQ=
 github.com/json-iterator/go v1.1.12 h1:PV8peI4a0ysnczrg+LtxykD8LfKY9ML6u2jnxaEnrnM=
 github.com/json-iterator/go v1.1.12/go.mod h1:e30LSqwooZae/UwlEbR2852Gd8hjQvJoHmT4TnhNGBo=
 github.com/klauspost/cpuid/v2 v2.0.9/go.mod h1:FInQzS24/EEf25PyTYn52gqo7WaD8xa0213Md/qVLRg=
@@ -64,14 +66,22 @@ golang.org/x/arch v0.8.0 h1:3wRIsP3pM4yUptoR96otTUOXI367OS0+c9eeRi9doIc=
 golang.org/x/arch v0.8.0/go.mod h1:FEVrYAQjsQXMVJ1nsMoVVXPZg6p2JE2mx8psSWTDQys=
 golang.org/x/crypto v0.23.0 h1:dIJU/v2J8Mdglj/8rJ6UUOM3Zc9zLZxVZwwxMooUSAI=
 golang.org/x/crypto v0.23.0/go.mod h1:CKFgDieR+mRhux2Lsu27y0fO304Db0wZe70UKqHu0v8=
+golang.org/x/crypto v0.31.0 h1:ihbySMvVjLAeSH1IbfcRTkD/iNscyz8rGzjF/E5hV6U=
+golang.org/x/crypto v0.31.0/go.mod h1:kDsLvtWBEx7MV9tJOj9bnXsPbxwJQ6csT/x4KIN4Ssk=
 golang.org/x/net v0.25.0 h1:d/OCCoBEUq33pjydKrGQhw7IlUPI2Oylr+8qLx49kac=
 golang.org/x/net v0.25.0/go.mod h1:JkAGAh7GEvH74S6FOH42FLoXpXbE/aqXSrIQjXgsiwM=
+golang.org/x/net v0.33.0 h1:74SYHlV8BIgHIFC/LrYkOGIwL19eTYXQ5wc6TBuO36I=
+golang.org/x/net v0.33.0/go.mod h1:HXLR5J+9DxmrqMwG9qjGCxZ+zKXxBru04zlTvWlWuN4=
 golang.org/x/sys v0.5.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.6.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.20.0 h1:Od9JTbYCk261bKm4M/mw7AklTlFYIa0bIp9BgSm1S8Y=
 golang.org/x/sys v0.20.0/go.mod h1:/VUhepiaJMQUp4+oa/7Zr1D23ma6VTLIYjOOTFZPUcA=
+golang.org/x/sys v0.28.0 h1:Fksou7UEQUWlKvIdsqzJmUmCX3cZuD2+P3XyyzwMhlA=
+golang.org/x/sys v0.28.0/go.mod h1:/VUhepiaJMQUp4+oa/7Zr1D23ma6VTLIYjOOTFZPUcA=
 golang.org/x/text v0.15.0 h1:h1V/4gjBv8v9cjcR6+AR5+/cIYK5N/WAgiv4xlsEtAk=
 golang.org/x/text v0.15.0/go.mod h1:18ZOQIKpY8NJVqYksKHtTdi31H5itFRjB5/qKTNYzSU=
+golang.org/x/text v0.21.0 h1:zyQAAkrwaneQ066sspRyJaG9VNi/YJ1NfzcGB3hZ/qo=
+golang.org/x/text v0.21.0/go.mod h1:4IBbMaMmOPCJ8SecivzSH54+73PCFmPWxNTLm+vZkEQ=
 google.golang.org/protobuf v1.34.1 h1:9ddQBjfCyZPOHPUiPxpYESBLc+T8P3E+Vo4IbKZgFWg=
 google.golang.org/protobuf v1.34.1/go.mod h1:c6P6GXX6sHbq/GpV6MGZEdwhWPcYBgnhAHhKbcUYpos=
 gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405 h1:yhCVgyC4o1eVCa2tZl7eS0r+SDo693bJlVdllGtEeKM=
@@ -4,18 +4,30 @@ import (
 	"flag"
 	"fmt"
 	"os"
+	"os/signal"
+	"syscall"

 	"github.com/gin-gonic/gin"
 	"github.com/mostlygeek/llama-swap/proxy"
 )

+var version string = "0"
+var commit string = "abcd1234"
+var date = "unknown"
+
 func main() {
 	// Define a command-line flag for the port
 	configPath := flag.String("config", "config.yaml", "config file name")
 	listenStr := flag.String("listen", ":8080", "listen ip/port")
+	showVersion := flag.Bool("version", false, "show version of build")

 	flag.Parse() // Parse the command-line flags

+	if *showVersion {
+		fmt.Printf("version: %s (%s), built at %s\n", version, commit, date)
+		os.Exit(0)
+	}
+
 	config, err := proxy.LoadConfig(*configPath)
 	if err != nil {
 		fmt.Printf("Error loading config: %v\n", err)
@@ -29,6 +41,16 @@ func main() {
 	}

 	proxyManager := proxy.New(config)
+
+	sigChan := make(chan os.Signal, 1)
+	signal.Notify(sigChan, syscall.SIGINT, syscall.SIGTERM)
+	go func() {
+		<-sigChan
+		fmt.Println("Shutting down llama-swap")
+		proxyManager.StopProcesses()
+		os.Exit(0)
+	}()
+
 	fmt.Println("llama-swap listening on " + *listenStr)
 	if err := proxyManager.Run(*listenStr); err != nil {
 		fmt.Printf("Server error: %v\n", err)
@@ -3,60 +3,137 @@ package main
 import (
 	"flag"
 	"fmt"
+	"io"
+	"log"
 	"net/http"
 	"os"
+	"os/signal"
+	"syscall"
+	"time"
+
+	"github.com/gin-gonic/gin"
 )

 func main() {
+	gin.SetMode(gin.TestMode)
 	// Define a command-line flag for the port
 	port := flag.String("port", "8080", "port to listen on")

 	// Define a command-line flag for the response message
 	responseMessage := flag.String("respond", "hi", "message to respond with")

+	silent := flag.Bool("silent", false, "disable all logging")
+
 	flag.Parse() // Parse the command-line flags

-	responseMessageHandler := func(w http.ResponseWriter, r *http.Request) {
-		// Set the header to text/plain
-		w.Header().Set("Content-Type", "text/plain")
-		fmt.Fprintln(w, *responseMessage)
-	}
+	// Create a new Gin router
+	r := gin.New()

 	// Set up the handler function using the provided response message
-	http.HandleFunc("/v1/chat/completions", responseMessageHandler)
-	http.HandleFunc("/v1/completions", responseMessageHandler)
-	http.HandleFunc("/test", responseMessageHandler)
+	r.POST("/v1/chat/completions", func(c *gin.Context) {
+		c.Header("Content-Type", "text/plain")

-	http.HandleFunc("/env", func(w http.ResponseWriter, r *http.Request) {
-		w.Header().Set("Content-Type", "text/plain")
-		fmt.Fprintln(w, *responseMessage)
+		// add a wait to simulate a slow query
+		if wait, err := time.ParseDuration(c.Query("wait")); err == nil {
+			time.Sleep(wait)
+		}
+
+		c.String(200, *responseMessage)
+	})
+
+	r.POST("/v1/completions", func(c *gin.Context) {
+		c.Header("Content-Type", "text/plain")
+		c.String(200, *responseMessage)
+	})
+
+	r.GET("/slow-respond", func(c *gin.Context) {
+		echo := c.Query("echo")
+		delay := c.Query("delay")
+
+		if echo == "" {
+			echo = "abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789"
+		}
+
+		// Parse the duration
+		if delay == "" {
+			delay = "100ms"
+		}
+
+		t, err := time.ParseDuration(delay)
+		if err != nil {
+			c.Header("Content-Type", "text/plain")
+			c.String(http.StatusBadRequest, fmt.Sprintf("Invalid duration: %s", err))
+			return
+		}
+
+		c.Header("Content-Type", "text/plain")
+		for _, char := range echo {
+			c.Writer.Write([]byte(string(char)))
+			c.Writer.Flush()
+
+			// wait
+			<-time.After(t)
+		}
+	})
+
+	r.GET("/test", func(c *gin.Context) {
+		c.Header("Content-Type", "text/plain")
+		c.String(200, *responseMessage)
+	})
+
+	r.GET("/env", func(c *gin.Context) {
+		c.Header("Content-Type", "text/plain")
+		c.String(200, *responseMessage)

 		// Get environment variables
 		envVars := os.Environ()

 		// Write each environment variable to the response
 		for _, envVar := range envVars {
-			fmt.Fprintln(w, envVar)
+			c.String(200, envVar)
 		}
 	})

 	// Set up the /health endpoint handler function
-	http.HandleFunc("/health", func(w http.ResponseWriter, r *http.Request) {
-		w.Header().Set("Content-Type", "application/json")
-		response := `{"status": "ok"}`
-		w.Write([]byte(response))
+	r.GET("/health", func(c *gin.Context) {
+		c.Header("Content-Type", "application/json")
+		c.JSON(200, gin.H{"status": "ok"})
 	})

-	http.HandleFunc("/", func(w http.ResponseWriter, r *http.Request) {
-		w.Header().Set("Content-Type", "text/plain")
-		fmt.Fprintf(w, "%s %s", r.Method, r.URL.Path)
+	r.GET("/", func(c *gin.Context) {
+		c.Header("Content-Type", "text/plain")
+		c.String(200, fmt.Sprintf("%s %s", c.Request.Method, c.Request.URL.Path))
 	})

 	address := "127.0.0.1:" + *port // Address with the specified port
-	fmt.Printf("Server is listening on port %s\n", *port)

-	// Start the server and log any error if it occurs
-	if err := http.ListenAndServe(address, nil); err != nil {
-		fmt.Printf("Error starting server: %s\n", err)
+	srv := &http.Server{
+		Addr:    address,
+		Handler: r.Handler(),
 	}
+
+	// Disable logging if the --silent flag is set
+	if *silent {
+		gin.SetMode(gin.ReleaseMode)
+		gin.DefaultWriter = io.Discard
+		log.SetOutput(io.Discard)
+	}
+
+	go func() {
+		log.Printf("simple-responder listening on %s\n", address)
+		// service connections
+		if err := srv.ListenAndServe(); err != nil && err != http.ErrServerClosed {
+			log.Fatalf("simple-responder err: %s\n", err)
+		}
+	}()
+
+	// Wait for interrupt signal to gracefully shutdown the server with
+	// a timeout of 5 seconds.
+	quit := make(chan os.Signal, 1)
+	// kill (no param) default send syscall.SIGTERM
+	// kill -2 is syscall.SIGINT
+	// kill -9 is syscall.SIGKILL but can't be catch, so don't need add it
+	signal.Notify(quit, syscall.SIGINT, syscall.SIGTERM)
+	<-quit
+	log.Println("simple-responder shutting down")
 }
@@ -0,0 +1,4 @@
+The rerank-test.json data is from https://github.com/ggerganov/llama.cpp/pull/9510
+
+To run it:
+> curl http://127.0.0.1:8080/v1/rerank -H "Content-Type: application/json" -d @reranker-test.json  -v | jq .
@@ -0,0 +1,17 @@
+{
+  "model": "bge-reranker",
+  "query": "Organic skincare products for sensitive skin",
+  "top_n": 3,
+  "documents": [
+    "Organic skincare for sensitive skin with aloe vera and chamomile: Imagine the soothing embrace of nature with our organic skincare range, crafted specifically for sensitive skin. Infused with the calming properties of aloe vera and chamomile, each product provides gentle nourishment and protection. Say goodbye to irritation and hello to a glowing, healthy complexion.",
+    "New makeup trends focus on bold colors and innovative techniques: Step into the world of cutting-edge beauty with this seasons makeup trends. Bold, vibrant colors and groundbreaking techniques are redefining the art of makeup. From neon eyeliners to holographic highlighters, unleash your creativity and make a statement with every look.",
+    "Bio-Hautpflege für empfindliche Haut mit Aloe Vera und Kamille: Erleben Sie die wohltuende Wirkung unserer Bio-Hautpflege, speziell für empfindliche Haut entwickelt. Mit den beruhigenden Eigenschaften von Aloe Vera und Kamille pflegen und schützen unsere Produkte Ihre Haut auf natürliche Weise. Verabschieden Sie sich von Hautirritationen und genießen Sie einen strahlenden Teint.",
+    "Neue Make-up-Trends setzen auf kräftige Farben und innovative Techniken: Tauchen Sie ein in die Welt der modernen Schönheit mit den neuesten Make-up-Trends. Kräftige, lebendige Farben und innovative Techniken setzen neue Maßstäbe. Von auffälligen Eyelinern bis hin zu holografischen Highlightern – lassen Sie Ihrer Kreativität freien Lauf und setzen Sie jedes Mal ein Statement.",
+    "Cuidado de la piel orgánico para piel sensible con aloe vera y manzanilla: Descubre el poder de la naturaleza con nuestra línea de cuidado de la piel orgánico, diseñada especialmente para pieles sensibles. Enriquecidos con aloe vera y manzanilla, estos productos ofrecen una hidratación y protección suave. Despídete de las irritaciones y saluda a una piel radiante y saludable.",
+    "Las nuevas tendencias de maquillaje se centran en colores vivos y técnicas innovadoras: Entra en el fascinante mundo del maquillaje con las tendencias más actuales. Colores vivos y técnicas innovadoras están revolucionando el arte del maquillaje. Desde delineadores neón hasta iluminadores holográficos, desata tu creatividad y destaca en cada look.",
+    "针对敏感肌专门设计的天然有机护肤产品：体验由芦荟和洋甘菊提取物带来的自然呵护。我们的护肤产品特别为敏感肌设计，温和滋润，保护您的肌肤不受刺激。让您的肌肤告别不适，迎来健康光彩。",
+    "新的化妆趋势注重鲜艳的颜色和创新的技巧：进入化妆艺术的新纪元，本季的化妆趋势以大胆的颜色和创新的技巧为主。无论是霓虹眼线还是全息高光，每一款妆容都能让您脱颖而出，展现独特魅力。",
+    "敏感肌のために特別に設計された天然有機スキンケア製品: アロエベラとカモミールのやさしい力で、自然の抱擁を感じてください。敏感肌用に特別に設計された私たちのスキンケア製品は、肌に優しく栄養を与え、保護します。肌トラブルにさようなら、輝く健康な肌にこんにちは。",
+    "新しいメイクのトレンドは鮮やかな色と革新的な技術に焦点を当てています: 今シーズンのメイクアップトレンドは、大胆な色彩と革新的な技術に注目しています。ネオンアイライナーからホログラフィックハイライターまで、クリエイティビティを解き放ち、毎回ユニークなルックを演出しましょう。"
+  ]
+}
@@ -5,6 +5,7 @@ import (
 	"os"
 	"strings"

+	"github.com/google/shlex"
 	"gopkg.in/yaml.v3"
 )

@@ -15,6 +16,7 @@ type ModelConfig struct {
 	Env           []string `yaml:"env"`
 	CheckEndpoint string   `yaml:"checkEndpoint"`
 	UnloadAfter   int      `yaml:"ttl"`
+	Unlisted      bool     `yaml:"unlisted"`
 }

 func (m *ModelConfig) SanitizedCommand() ([]string, error) {
@@ -23,6 +25,7 @@ func (m *ModelConfig) SanitizedCommand() ([]string, error) {

 type Config struct {
 	HealthCheckTimeout int                    `yaml:"healthCheckTimeout"`
+	LogRequests        bool                   `yaml:"logRequests"`
 	Models             map[string]ModelConfig `yaml:"models"`
 	Profiles           map[string][]string    `yaml:"profiles"`

@@ -81,7 +84,10 @@ func SanitizeCommand(cmdStr string) ([]string, error) {
 	cmdStr = strings.ReplaceAll(cmdStr, "\\\n", " ")

 	// Split the command into arguments
-	args := strings.Fields(cmdStr)
+	args, err := shlex.Split(cmdStr)
+	if err != nil {
+		return nil, err
+	}

 	// Ensure the command is not empty
 	if len(args) == 0 {
@@ -148,17 +148,26 @@ func TestConfig_FindConfig(t *testing.T) {
 }

 func TestConfig_SanitizeCommand(t *testing.T) {
-	// Test a simple command
-	args, err := SanitizeCommand("python model1.py")
-	assert.NoError(t, err)
-	assert.Equal(t, []string{"python", "model1.py"}, args)

 	// Test a command with spaces and newlines
-	args, err = SanitizeCommand(`python model1.py \
-    --arg1 value1 \
-    --arg2 value2`)
+	args, err := SanitizeCommand(`python model1.py \
+    -a "double quotes" \
+    --arg2 'single quotes'
+	-s
+	--arg3 123 \
+	--arg4 '"string in string"'
+	-c "'single quoted'"
+	`)
 	assert.NoError(t, err)
-	assert.Equal(t, []string{"python", "model1.py", "--arg1", "value1", "--arg2", "value2"}, args)
+	assert.Equal(t, []string{
+		"python", "model1.py",
+		"-a", "double quotes",
+		"--arg2", "single quotes",
+		"-s",
+		"--arg3", "123",
+		"--arg4", `"string in string"`,
+		"-c", `'single quoted'`,
+	}, args)

 	// Test an empty command
 	args, err = SanitizeCommand("")
@@ -51,7 +51,7 @@ func getTestSimpleResponderConfigPort(expectedMessage string, port int) ModelCon

 	// Create a process configuration
 	return ModelConfig{
-		Cmd:           fmt.Sprintf("%s --port %d --respond '%s'", binaryPath, port, expectedMessage),
+		Cmd:           fmt.Sprintf("%s --port %d --silent --respond %s", binaryPath, port, expectedMessage),
 		Proxy:         fmt.Sprintf("http://127.0.0.1:%d", port),
 		CheckEndpoint: "/health",
 	}
@@ -0,0 +1,14 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+    <meta charset="UTF-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1.0">
+    <title>llama-swap</title>
+</head>
+<body>
+    <h1>llama-swap</h1>
+    <p>
+        <a href="/logs">view logs</a> | <a href="/upstream">configured models</a> | <a href="https://github.com/mostlygeek/llama-swap">github</a>
+    </p>
+</body>
+</html>
@@ -0,0 +1,145 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+    <meta charset="UTF-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1.0">
+    <title>Logs</title>
+    <style>
+        body {
+            margin: 0;
+            height: 100vh;
+            display: flex;
+            flex-direction: column;
+            font-family: "Courier New", Courier, monospace;
+        }
+        #log-controls {
+            margin: 0.5em;
+            display: flex;
+            align-items: center;
+            justify-content: space-between; /* Spaces out elements evenly */
+        }
+        #log-controls input {
+            flex: 1;
+        }
+        #log-controls input:focus {
+           outline: none; /* Ensures no outline is shown when the input is focused */
+        }
+        #log-stream {
+            flex: 1;
+            margin: 0.5em;
+            padding: 1em;
+            background: #f4f4f4;
+            overflow-y: auto;
+            white-space: pre-wrap; /* Ensures line wrapping */
+            word-wrap: break-word; /* Ensures long words wrap */
+        }
+
+        .regex-error {
+            background-color: #ff0000 !important;
+        }
+
+        /* Dark mode styles */
+        @media (prefers-color-scheme: dark) {
+            body {
+                background-color: #333;
+                color: #fff;
+            }
+
+            #log-stream {
+                background: #444;
+                color: #fff;
+            }
+
+            #log-controls input {
+                background: #555;
+                color: #fff;
+                border: 1px solid #777;
+            }
+
+            #log-controls button {
+                background: #555;
+                color: #fff;
+                border: 1px solid #777;
+            }
+        }
+    </style>
+</head>
+<body>
+    <pre id="log-stream">Waiting for logs...</pre>
+    <div id="log-controls">
+        <input type="text" id="filter-input" placeholder="regex filter">
+        <button id="clear-button">clear</button>
+    </div>
+    <script>
+        const logStream = document.getElementById('log-stream');
+        const filterInput = document.getElementById('filter-input');
+        var logData = "";
+        let regexFilter = null;
+
+        function setupEventSource() {
+            if (typeof(EventSource) !== "undefined") {
+                const eventSource = new EventSource("/logs/streamSSE");
+
+                eventSource.onmessage = function(event) {
+                    logData += event.data;
+                    render()
+                };
+
+                eventSource.onerror = function(err) {
+                    logData = "EventSource failed: " + err.message;
+                };
+            } else {
+                logData = "SSE Not supported by this browser."
+            }
+        }
+
+        // poor-ai's react  ¯\_(ツ)_/¯
+        function render() {
+            if (regexFilter) {
+                const lines = logData.split('\n');
+                const filteredLines = lines.filter(line => {
+                    return regexFilter === null || regexFilter.test(line);
+                });
+
+                if (filteredLines.length > 0) {
+                    logStream.textContent = filteredLines.join('\n') + '\n';
+                } else {
+                    logStream.textContent = "";
+                }
+            } else {
+                logStream.textContent = logData;
+            }
+
+            logStream.scrollTop = logStream.scrollHeight;
+        }
+
+        function updateFilter() {
+            const pattern = filterInput.value.trim();
+            filterInput.classList.remove('regex-error');
+            if (pattern) {
+                try {
+                    regexFilter = new RegExp(pattern);
+                } catch (e) {
+                    console.error("Invalid regex pattern:", e);
+                    regexFilter = null;
+                    filterInput.classList.add('regex-error');
+                    return
+                }
+            } else {
+                regexFilter = null;
+            }
+
+            render();
+        }
+
+        filterInput.addEventListener('input', updateFilter);
+        document.getElementById('clear-button').addEventListener('click', () => {
+            filterInput.value = "";
+            regexFilter = null;
+            render();
+        });
+        setupEventSource();
+        updateFilter();
+    </script>
+</body>
+</html>
@@ -0,0 +1,10 @@
+package proxy
+
+import "embed"
+
+//go:embed html
+var htmlFiles embed.FS
+
+func getHTMLFile(path string) ([]byte, error) {
+	return htmlFiles.ReadFile("html/" + path)
+}
@@ -46,7 +46,7 @@ func (w *LogMonitor) Write(p []byte) (n int, err error) {
 	w.buffer = w.buffer.Next()
 	w.bufferMu.Unlock()

-	w.broadcast(p)
+	w.broadcast(bufferCopy)
 	return n, nil
 }

@@ -14,6 +14,14 @@ import (
 	"time"
 )

+type ProcessState string
+
+const (
+	StateStopped ProcessState = ProcessState("stopped")
+	StateReady   ProcessState = ProcessState("ready")
+	StateFailed  ProcessState = ProcessState("failed")
+)
+
 type Process struct {
 	sync.Mutex

@@ -23,8 +31,12 @@ type Process struct {
 	logMonitor         *LogMonitor
 	healthCheckTimeout int

-	isRunning          bool
 	lastRequestHandled time.Time
+
+	stateMutex sync.RWMutex
+	state      ProcessState
+
+	inFlightRequests sync.WaitGroup
 }

 func NewProcess(ID string, healthCheckTimeout int, config ModelConfig, logMonitor *LogMonitor) *Process {
@@ -34,16 +46,22 @@ func NewProcess(ID string, healthCheckTimeout int, config ModelConfig, logMonito
 		cmd:                nil,
 		logMonitor:         logMonitor,
 		healthCheckTimeout: healthCheckTimeout,
+		state:              StateStopped,
 	}
 }

-// start the process and check it for errors
+// start the process and returns when it is ready
 func (p *Process) start() error {
-	p.Lock()
-	defer p.Unlock()

-	if p.isRunning {
-		return fmt.Errorf("process already running")
+	p.stateMutex.Lock()
+	defer p.stateMutex.Unlock()
+
+	if p.state == StateReady {
+		return nil
+	}
+
+	if p.state == StateFailed {
+		return fmt.Errorf("process is in a failed state and can not be restarted")
 	}

 	args, err := p.config.SanitizedCommand()
@@ -57,48 +75,65 @@ func (p *Process) start() error {
 	p.cmd.Env = p.config.Env

 	err = p.cmd.Start()
-	p.isRunning = true

 	if err != nil {
 		return err
 	}

-	// watch for the command to exit
-	cmdCtx, cancel := context.WithCancelCause(context.Background())
+	// One of three things can happen at this stage:
+	// 1. The command exits unexpectedly
+	// 2. The health check fails
+	// 3. The health check passes
+	//
+	// only in the third case will the process be considered Ready to accept
+	healthCheckContext, cancelHealthCheck := context.WithCancelCause(context.Background())
+	defer cancelHealthCheck(nil) // clean up
+	cmdWaitChan := make(chan error, 1)
+	healthCheckChan := make(chan error, 1)

-	// monitor the command's exit status. Usually this happens if
-	// the process exited unexpectedly
 	go func() {
-		err := p.cmd.Wait()
-		if err != nil {
-			cancel(fmt.Errorf("command [%s] %s", strings.Join(p.cmd.Args, " "), err.Error()))
-		} else {
-			cancel(nil)
-		}
-
-		p.isRunning = false
+		// possible cmd exits early
+		cmdWaitChan <- p.cmd.Wait()
 	}()

-	// wait a bit for process to start before checking the health endpoint
-	time.Sleep(250 * time.Millisecond)
+	go func() {
+		<-time.After(250 * time.Millisecond) // give process a bit of time to start
+		healthCheckChan <- p.checkHealthEndpoint(healthCheckContext)
+	}()

-	// wait for checkHealthEndpoint
-	if err := p.checkHealthEndpoint(cmdCtx); err != nil {
+	select {
+	case err := <-cmdWaitChan:
+		p.state = StateFailed
+		if err != nil {
+			err = fmt.Errorf("command [%s] %s", strings.Join(p.cmd.Args, " "), err.Error())
+		} else {
+			err = fmt.Errorf("command [%s] exited unexpected", strings.Join(p.cmd.Args, " "))
+		}
+		cancelHealthCheck(err)
 		return err
+	case err := <-healthCheckChan:
+		if err != nil {
+			p.state = StateFailed
+			return err
+		}
 	}

 	if p.config.UnloadAfter > 0 {
 		// start a goroutine to check every second if
 		// the process should be stopped
 		go func() {
-			ticker := time.NewTicker(time.Second)
-			defer ticker.Stop()
 			maxDuration := time.Duration(p.config.UnloadAfter) * time.Second

-			for {
-				<-ticker.C
+			for range time.Tick(time.Second) {
+				if p.state != StateReady {
+					return
+				}
+
+				// wait for all inflight requests to complete and ticker
+				p.inFlightRequests.Wait()
+
 				if time.Since(p.lastRequestHandled) > maxDuration {
-					fmt.Fprintf(p.logMonitor, "!!! Unloading model %s, TTL of %d reached.\n", p.ID, p.config.UnloadAfter)
+					fmt.Fprintf(p.logMonitor, "!!! Unloading model %s, TTL of %ds reached.\n", p.ID, p.config.UnloadAfter)
 					p.Stop()
 					return
 				}
@@ -106,27 +141,64 @@ func (p *Process) start() error {
 		}()
 	}

+	p.state = StateReady
 	return nil
 }

 func (p *Process) Stop() {
-	p.Lock()
-	defer p.Unlock()
+	// wait for any inflight requests before proceeding
+	p.inFlightRequests.Wait()

-	if !p.isRunning || p.cmd == nil || p.cmd.Process == nil {
+	p.stateMutex.Lock()
+	defer p.stateMutex.Unlock()
+
+	if p.state != StateReady {
+		fmt.Fprintf(p.logMonitor, "!!! Info - Stop() called but Process State is not READY\n")
 		return
 	}

+	if p.cmd == nil || p.cmd.Process == nil {
+		// this situation should never happen... but if it does just update the state
+		fmt.Fprintf(p.logMonitor, "!!! State is Ready but Command is nil.\n")
+		p.state = StateStopped
+		return
+	}
+
+	sigtermTimeout, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+
+	sigtermNormal := make(chan error, 1)
+	go func() {
+		sigtermNormal <- p.cmd.Wait()
+	}()
+
 	p.cmd.Process.Signal(syscall.SIGTERM)
-	p.cmd.Process.Wait()
-	p.isRunning = false
+
+	select {
+	case <-sigtermTimeout.Done():
+		fmt.Fprintf(p.logMonitor, "XXX Process for %s timed out waiting to stop, sending SIGKILL to PID: %d\n", p.ID, p.cmd.Process.Pid)
+		p.cmd.Process.Kill()
+		p.cmd.Wait()
+	case err := <-sigtermNormal:
+		if err != nil {
+			if err.Error() != "wait: no child processes" {
+				// possible that simple-responder for testing is just not
+				// existing right, so suppress those errors.
+				fmt.Fprintf(p.logMonitor, "!!! process for %s stopped with error > %v\n", p.ID, err)
+			}
+		}
+	}
+
+	p.state = StateStopped
 }

-func (p *Process) IsRunning() bool {
-	return p.isRunning
+func (p *Process) CurrentState() ProcessState {
+	p.stateMutex.RLock()
+	defer p.stateMutex.RUnlock()
+	return p.state
 }

-func (p *Process) checkHealthEndpoint(cmdCtx context.Context) error {
+func (p *Process) checkHealthEndpoint(ctxFromStart context.Context) error {
 	if p.config.Proxy == "" {
 		return fmt.Errorf("no upstream available to check /health")
 	}
@@ -158,7 +230,7 @@ func (p *Process) checkHealthEndpoint(cmdCtx context.Context) error {
 			return err
 		}

-		ctx, cancel := context.WithTimeout(cmdCtx, time.Second)
+		ctx, cancel := context.WithTimeout(ctxFromStart, time.Second)
 		defer cancel()
 		req = req.WithContext(ctx)
 		resp, err := client.Do(req)
@@ -205,7 +277,15 @@ func (p *Process) checkHealthEndpoint(cmdCtx context.Context) error {
 }

 func (p *Process) ProxyRequest(w http.ResponseWriter, r *http.Request) {
-	if !p.isRunning {
+
+	p.inFlightRequests.Add(1)
+
+	defer func() {
+		p.lastRequestHandled = time.Now()
+		p.inFlightRequests.Done()
+	}()
+
+	if p.CurrentState() != StateReady {
 		if err := p.start(); err != nil {
 			errstr := fmt.Sprintf("unable to start process: %s", err)
 			http.Error(w, errstr, http.StatusInternalServerError)
@@ -213,11 +293,9 @@ func (p *Process) ProxyRequest(w http.ResponseWriter, r *http.Request) {
 		}
 	}

-	p.lastRequestHandled = time.Now()
-
 	proxyTo := p.config.Proxy
 	client := &http.Client{}
-	req, err := http.NewRequest(r.Method, proxyTo+r.URL.String(), r.Body)
+	req, err := http.NewRequestWithContext(r.Context(), r.Method, proxyTo+r.URL.String(), r.Body)
 	if err != nil {
 		http.Error(w, err.Error(), http.StatusInternalServerError)
 		return
@@ -1,9 +1,12 @@
 package proxy

 import (
+	"fmt"
 	"io"
 	"net/http"
 	"net/http/httptest"
+	"os"
+	"sync"
 	"testing"
 	"time"

@@ -23,9 +26,9 @@ func TestProcess_AutomaticallyStartsUpstream(t *testing.T) {
 	w := httptest.NewRecorder()

 	// process is automatically started
-	assert.False(t, process.IsRunning())
+	assert.Equal(t, StateStopped, process.CurrentState())
 	process.ProxyRequest(w, req)
-	assert.True(t, process.IsRunning())
+	assert.Equal(t, StateReady, process.CurrentState())

 	assert.Equal(t, http.StatusOK, w.Code, "Expected status code %d, got %d", http.StatusOK, w.Code)
 	assert.Contains(t, w.Body.String(), expectedMessage)
@@ -49,7 +52,7 @@ func TestProcess_AutomaticallyStartsUpstream(t *testing.T) {
 func TestProcess_BrokenModelConfig(t *testing.T) {
 	// Create a process configuration
 	config := ModelConfig{
-		Cmd:           "nonexistant-command",
+		Cmd:           "nonexistent-command",
 		Proxy:         "http://127.0.0.1:9913",
 		CheckEndpoint: "/health",
 	}
@@ -64,7 +67,6 @@ func TestProcess_BrokenModelConfig(t *testing.T) {
 	assert.Contains(t, w.Body.String(), "unable to start process")
 }

-// test that the process unloads after the TTL
 func TestProcess_UnloadAfterTTL(t *testing.T) {
 	if testing.Short() {
 		t.Skip("skipping long auto unload TTL test")
@@ -76,21 +78,115 @@ func TestProcess_UnloadAfterTTL(t *testing.T) {
 	config.UnloadAfter = 3 // seconds
 	assert.Equal(t, 3, config.UnloadAfter)

-	process := NewProcess("ttl", 2, config, NewLogMonitorWriter(io.Discard))
+	process := NewProcess("ttl_test", 2, config, NewLogMonitorWriter(io.Discard))
 	defer process.Stop()

-	req := httptest.NewRequest("GET", "/test", nil)
+	// this should take 4 seconds
+	req1 := httptest.NewRequest("GET", "/slow-respond?echo=1234&delay=1000ms", nil)
+	req2 := httptest.NewRequest("GET", "/test", nil)
+
 	w := httptest.NewRecorder()

-	// Proxy the request (auto start)
-	process.ProxyRequest(w, req)
+	// Proxy the request (auto start) with a slow response that takes longer than config.UnloadAfter
+	process.ProxyRequest(w, req1)
+
+	t.Log("sending slow first request (4 seconds)")
+	assert.Equal(t, http.StatusOK, w.Code, "Expected status code %d, got %d", http.StatusOK, w.Code)
+	assert.Contains(t, w.Body.String(), "1234")
+	assert.Equal(t, StateReady, process.CurrentState())
+
+	// ensure the TTL timeout does not race slow requests (see issue #25)
+	t.Log("sending second request (1 second)")
+	time.Sleep(time.Second)
+	w = httptest.NewRecorder()
+	process.ProxyRequest(w, req2)
 	assert.Equal(t, http.StatusOK, w.Code, "Expected status code %d, got %d", http.StatusOK, w.Code)
 	assert.Contains(t, w.Body.String(), expectedMessage)
-
-	assert.True(t, process.IsRunning())
+	assert.Equal(t, StateReady, process.CurrentState())

 	// wait 5 seconds
+	t.Log("sleep 5 seconds and check if unloaded")
 	time.Sleep(5 * time.Second)
-
-	assert.False(t, process.IsRunning())
+	assert.Equal(t, StateStopped, process.CurrentState())
+}
+
+func TestProcess_LowTTLValue(t *testing.T) {
+	if true { // change this code to run this ...
+		t.Skip("skipping test, edit process_test.go to run it ")
+	}
+
+	config := getTestSimpleResponderConfig("fast_ttl")
+	assert.Equal(t, 0, config.UnloadAfter)
+	config.UnloadAfter = 1 // second
+	assert.Equal(t, 1, config.UnloadAfter)
+
+	process := NewProcess("ttl", 2, config, NewLogMonitorWriter(os.Stdout))
+	defer process.Stop()
+
+	for i := 0; i < 100; i++ {
+		t.Logf("Waiting before sending request %d", i)
+		time.Sleep(1500 * time.Millisecond)
+
+		expected := fmt.Sprintf("echo=test_%d", i)
+		req := httptest.NewRequest("GET", fmt.Sprintf("/slow-respond?echo=%s&delay=50ms", expected), nil)
+		w := httptest.NewRecorder()
+		process.ProxyRequest(w, req)
+		assert.Equal(t, http.StatusOK, w.Code)
+		assert.Contains(t, w.Body.String(), expected)
+	}
+
+}
+
+// issue #19
+func TestProcess_HTTPRequestsHaveTimeToFinish(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping slow test")
+	}
+
+	expectedMessage := "12345"
+	config := getTestSimpleResponderConfig(expectedMessage)
+	process := NewProcess("t", 10, config, NewLogMonitorWriter(os.Stdout))
+	defer process.Stop()
+
+	results := map[string]string{
+		"12345": "",
+		"abcde": "",
+		"fghij": "",
+	}
+
+	var wg sync.WaitGroup
+	var mu sync.Mutex
+
+	for key := range results {
+		wg.Add(1)
+		go func(key string) {
+			defer wg.Done()
+			// send a request that should take 5 * 200ms (1 second) to complete
+			req := httptest.NewRequest("GET", fmt.Sprintf("/slow-respond?echo=%s&delay=200ms", key), nil)
+			w := httptest.NewRecorder()
+
+			process.ProxyRequest(w, req)
+
+			if w.Code != http.StatusOK {
+				t.Errorf("Expected status OK, got %d for key %s", w.Code, key)
+			}
+
+			mu.Lock()
+			results[key] = w.Body.String()
+			mu.Unlock()
+
+		}(key)
+	}
+
+	// stop the requests in the middle
+	go func() {
+		<-time.After(500 * time.Millisecond)
+		process.Stop()
+	}()
+
+	wg.Wait()
+
+	for key, result := range results {
+		assert.Equal(t, key, result)
+	}
 }
@@ -6,6 +6,7 @@ import (
 	"fmt"
 	"io"
 	"net/http"
+	"sort"
 	"strconv"
 	"strings"
 	"sync"
@@ -14,6 +15,10 @@ import (
 	"github.com/gin-gonic/gin"
 )

+const (
+	PROFILE_SPLIT_CHAR = ":"
+)
+
 type ProxyManager struct {
 	sync.Mutex

@@ -31,11 +36,63 @@ func New(config *Config) *ProxyManager {
 		ginEngine:        gin.New(),
 	}

-	// Set up routes using the Gin engine
-	pm.ginEngine.POST("/v1/chat/completions", pm.proxyChatRequestHandler)
+	if config.LogRequests {
+		pm.ginEngine.Use(func(c *gin.Context) {
+			// Start timer
+			start := time.Now()

+			// capture these because /upstream/:model rewrites them in c.Next()
+			clientIP := c.ClientIP()
+			method := c.Request.Method
+			path := c.Request.URL.Path
+
+			// Process request
+			c.Next()
+
+			// Stop timer
+			duration := time.Since(start)
+
+			statusCode := c.Writer.Status()
+			bodySize := c.Writer.Size()
+
+			fmt.Fprintf(pm.logMonitor, "[llama-swap] %s [%s] \"%s %s %s\" %d %d \"%s\" %v\n",
+				clientIP,
+				time.Now().Format("2006-01-02 15:04:05"),
+				method,
+				path,
+				c.Request.Proto,
+				statusCode,
+				bodySize,
+				c.Request.UserAgent(),
+				duration,
+			)
+		})
+	}
+
+	// see: https://github.com/mostlygeek/llama-swap/issues/42
+	// respond with permissive OPTIONS for any endpoint
+	pm.ginEngine.Use(func(c *gin.Context) {
+		if c.Request.Method == "OPTIONS" {
+			c.Header("Access-Control-Allow-Origin", "*")
+			c.Header("Access-Control-Allow-Methods", "GET, POST, OPTIONS")
+			c.Header("Access-Control-Allow-Headers", "Content-Type, Authorization")
+			c.AbortWithStatus(204)
+			return
+		}
+		c.Next()
+	})
+
+	// Set up routes using the Gin engine
+	pm.ginEngine.POST("/v1/chat/completions", pm.proxyOAIHandler)
 	// Support legacy /v1/completions api, see issue #12
-	pm.ginEngine.POST("/v1/completions", pm.proxyChatRequestHandler)
+	pm.ginEngine.POST("/v1/completions", pm.proxyOAIHandler)
+
+	// Support embeddings
+	pm.ginEngine.POST("/v1/embeddings", pm.proxyOAIHandler)
+	pm.ginEngine.POST("/v1/rerank", pm.proxyOAIHandler)
+
+	// Support audio/speech endpoint
+	pm.ginEngine.POST("/v1/audio/speech", pm.proxyOAIHandler)

 	pm.ginEngine.GET("/v1/models", pm.listModelsHandler)

@@ -44,7 +101,33 @@ func New(config *Config) *ProxyManager {
 	pm.ginEngine.GET("/logs/stream", pm.streamLogsHandler)
 	pm.ginEngine.GET("/logs/streamSSE", pm.streamLogsHandlerSSE)

-	pm.ginEngine.NoRoute(pm.proxyNoRouteHandler)
+	pm.ginEngine.GET("/upstream", pm.upstreamIndex)
+	pm.ginEngine.Any("/upstream/:model_id/*upstreamPath", pm.proxyToUpstream)
+
+	pm.ginEngine.GET("/", func(c *gin.Context) {
+		// Set the Content-Type header to text/html
+		c.Header("Content-Type", "text/html")
+
+		// Write the embedded HTML content to the response
+		htmlData, err := getHTMLFile("index.html")
+		if err != nil {
+			c.String(http.StatusInternalServerError, err.Error())
+			return
+		}
+		_, err = c.Writer.Write(htmlData)
+		if err != nil {
+			c.String(http.StatusInternalServerError, fmt.Sprintf("failed to write response: %v", err))
+			return
+		}
+	})
+
+	pm.ginEngine.GET("/favicon.ico", func(c *gin.Context) {
+		if data, err := getHTMLFile("favicon.ico"); err == nil {
+			c.Data(http.StatusOK, "image/x-icon", data)
+		} else {
+			c.String(http.StatusInternalServerError, err.Error())
+		}
+	})

 	// Disable console color for testing
 	gin.DisableConsoleColor()
@@ -69,15 +152,24 @@ func (pm *ProxyManager) StopProcesses() {

 // for internal usage
 func (pm *ProxyManager) stopProcesses() {
+	if len(pm.currentProcesses) == 0 {
+		return
+	}
+
 	for _, process := range pm.currentProcesses {
 		process.Stop()
 	}
+
 	pm.currentProcesses = make(map[string]*Process)
 }

 func (pm *ProxyManager) listModelsHandler(c *gin.Context) {
 	data := []interface{}{}
-	for id := range pm.config.Models {
+	for id, modelConfig := range pm.config.Models {
+		if modelConfig.Unlisted {
+			continue
+		}
+
 		data = append(data, map[string]interface{}{
 			"id":       id,
 			"object":   "model",
@@ -89,9 +181,13 @@ func (pm *ProxyManager) listModelsHandler(c *gin.Context) {
 	// Set the Content-Type header to application/json
 	c.Header("Content-Type", "application/json")

+	if origin := c.Request.Header.Get("Origin"); origin != "" {
+		c.Header("Access-Control-Allow-Origin", origin)
+	}
+
 	// Encode the data as JSON and write it to the response writer
 	if err := json.NewEncoder(c.Writer).Encode(map[string]interface{}{"data": data}); err != nil {
-		c.AbortWithError(http.StatusInternalServerError, fmt.Errorf("error encoding JSON"))
+		pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error encoding JSON %s", err.Error()))
 		return
 	}
 }
@@ -100,16 +196,16 @@ func (pm *ProxyManager) swapModel(requestedModel string) (*Process, error) {
 	pm.Lock()
 	defer pm.Unlock()

-	// Check if requestedModel contains a /
-	groupName, modelName := "", requestedModel
-	if idx := strings.Index(requestedModel, "/"); idx != -1 {
-		groupName = requestedModel[:idx]
+	// Check if requestedModel contains a PROFILE_SPLIT_CHAR
+	profileName, modelName := "", requestedModel
+	if idx := strings.Index(requestedModel, PROFILE_SPLIT_CHAR); idx != -1 {
+		profileName = requestedModel[:idx]
 		modelName = requestedModel[idx+1:]
 	}

-	if groupName != "" {
-		if _, found := pm.config.Profiles[groupName]; !found {
-			return nil, fmt.Errorf("model group not found %s", groupName)
+	if profileName != "" {
+		if _, found := pm.config.Profiles[profileName]; !found {
+			return nil, fmt.Errorf("model group not found %s", profileName)
 		}
 	}

@@ -119,8 +215,24 @@ func (pm *ProxyManager) swapModel(requestedModel string) (*Process, error) {
 		return nil, fmt.Errorf("could not find modelID for %s", requestedModel)
 	}

+	// check if model is part of the profile
+	if profileName != "" {
+		found := false
+		for _, item := range pm.config.Profiles[profileName] {
+			if item == realModelName {
+				found = true
+				break
+			}
+		}
+
+		if !found {
+			return nil, fmt.Errorf("model %s part of profile %s", realModelName, profileName)
+		}
+	}
+
 	// exit early when already running, otherwise stop everything and swap
-	requestedProcessKey := groupName + "/" + realModelName
+	requestedProcessKey := ProcessKeyName(profileName, realModelName)
+
 	if process, found := pm.currentProcesses[requestedProcessKey]; found {
 		return process, nil
 	}
@@ -128,25 +240,25 @@ func (pm *ProxyManager) swapModel(requestedModel string) (*Process, error) {
 	// stop all running models
 	pm.stopProcesses()

-	if groupName == "" {
+	if profileName == "" {
 		modelConfig, modelID, found := pm.config.FindConfig(realModelName)
 		if !found {
 			return nil, fmt.Errorf("could not find configuration for %s", realModelName)
 		}

 		process := NewProcess(modelID, pm.config.HealthCheckTimeout, modelConfig, pm.logMonitor)
-		processKey := groupName + "/" + modelID
+		processKey := ProcessKeyName(profileName, modelID)
 		pm.currentProcesses[processKey] = process
 	} else {
-		for _, modelName := range pm.config.Profiles[groupName] {
+		for _, modelName := range pm.config.Profiles[profileName] {
 			if realModelName, found := pm.config.RealModelName(modelName); found {
 				modelConfig, modelID, found := pm.config.FindConfig(realModelName)
 				if !found {
-					return nil, fmt.Errorf("could not find configuration for %s in group %s", realModelName, groupName)
+					return nil, fmt.Errorf("could not find configuration for %s in group %s", realModelName, profileName)
 				}

 				process := NewProcess(modelID, pm.config.HealthCheckTimeout, modelConfig, pm.logMonitor)
-				processKey := groupName + "/" + modelID
+				processKey := ProcessKeyName(profileName, modelID)
 				pm.currentProcesses[processKey] = process
 			}
 		}
@@ -156,25 +268,68 @@ func (pm *ProxyManager) swapModel(requestedModel string) (*Process, error) {
 	return pm.currentProcesses[requestedProcessKey], nil
 }

-func (pm *ProxyManager) proxyChatRequestHandler(c *gin.Context) {
-	bodyBytes, err := io.ReadAll(c.Request.Body)
-	if err != nil {
-		c.AbortWithError(http.StatusBadRequest, fmt.Errorf("invalid JSON"))
+func (pm *ProxyManager) proxyToUpstream(c *gin.Context) {
+	requestedModel := c.Param("model_id")
+
+	if requestedModel == "" {
+		pm.sendErrorResponse(c, http.StatusBadRequest, "model id required in path")
 		return
 	}
+
+	if process, err := pm.swapModel(requestedModel); err != nil {
+		pm.sendErrorResponse(c, http.StatusNotFound, fmt.Sprintf("unable to swap to model, %s", err.Error()))
+	} else {
+		// rewrite the path
+		c.Request.URL.Path = c.Param("upstreamPath")
+		process.ProxyRequest(c.Writer, c.Request)
+	}
+}
+
+func (pm *ProxyManager) upstreamIndex(c *gin.Context) {
+	var html strings.Builder
+
+	html.WriteString("<!doctype HTML>\n<html><body><h1>Available Models</h1><ul>")
+
+	// Extract keys and sort them
+	var modelIDs []string
+	for modelID, modelConfig := range pm.config.Models {
+		if modelConfig.Unlisted {
+			continue
+		}
+
+		modelIDs = append(modelIDs, modelID)
+	}
+	sort.Strings(modelIDs)
+
+	// Iterate over sorted keys
+	for _, modelID := range modelIDs {
+		html.WriteString(fmt.Sprintf("<li><a href=\"/upstream/%s\">%s</a></li>", modelID, modelID))
+	}
+	html.WriteString("</ul></body></html>")
+	c.Header("Content-Type", "text/html")
+	c.String(http.StatusOK, html.String())
+}
+
+func (pm *ProxyManager) proxyOAIHandler(c *gin.Context) {
+	bodyBytes, err := io.ReadAll(c.Request.Body)
+	if err != nil {
+		pm.sendErrorResponse(c, http.StatusBadRequest, "could not ready request body")
+		return
+	}
+
 	var requestBody map[string]interface{}
 	if err := json.Unmarshal(bodyBytes, &requestBody); err != nil {
-		c.AbortWithError(http.StatusBadRequest, fmt.Errorf("invalid JSON"))
+		pm.sendErrorResponse(c, http.StatusBadRequest, fmt.Sprintf("invalid JSON: %s", err.Error()))
 		return
 	}
 	model, ok := requestBody["model"].(string)
 	if !ok {
-		c.AbortWithError(http.StatusBadRequest, fmt.Errorf("missing or invalid 'model' key"))
+		pm.sendErrorResponse(c, http.StatusBadRequest, "missing or invalid 'model' key")
 		return
 	}

 	if process, err := pm.swapModel(model); err != nil {
-		c.AbortWithError(http.StatusNotFound, fmt.Errorf("unable to swap to model, %s", err.Error()))
+		pm.sendErrorResponse(c, http.StatusNotFound, fmt.Sprintf("unable to swap to model, %s", err.Error()))
 		return
 	} else {
 		c.Request.Body = io.NopCloser(bytes.NewBuffer(bodyBytes))
@@ -185,15 +340,18 @@ func (pm *ProxyManager) proxyChatRequestHandler(c *gin.Context) {

 		process.ProxyRequest(c.Writer, c.Request)
 	}
-
 }

-func (pm *ProxyManager) proxyNoRouteHandler(c *gin.Context) {
-	// since maps are unordered, just use the first available process if one exists
-	for _, process := range pm.currentProcesses {
-		process.ProxyRequest(c.Writer, c.Request)
-		return
+func (pm *ProxyManager) sendErrorResponse(c *gin.Context, statusCode int, message string) {
+	acceptHeader := c.GetHeader("Accept")
+
+	if strings.Contains(acceptHeader, "application/json") {
+		c.JSON(statusCode, gin.H{"error": message})
+	} else {
+		c.String(statusCode, message)
 	}
-
-	c.AbortWithError(http.StatusBadRequest, fmt.Errorf("no strategy to handle request"))
+}
+
+func ProcessKeyName(groupName, modelName string) string {
+	return groupName + PROFILE_SPLIT_CHAR + modelName
 }
@@ -3,17 +3,37 @@ package proxy
 import (
 	"fmt"
 	"net/http"
+	"strings"

 	"github.com/gin-gonic/gin"
 )

 func (pm *ProxyManager) sendLogsHandlers(c *gin.Context) {
-	c.Header("Content-Type", "text/plain")
-	history := pm.logMonitor.GetHistory()
-	_, err := c.Writer.Write(history)
-	if err != nil {
-		c.AbortWithError(http.StatusInternalServerError, err)
-		return
+
+	accept := c.GetHeader("Accept")
+	if strings.Contains(accept, "text/html") {
+		// Set the Content-Type header to text/html
+		c.Header("Content-Type", "text/html")
+
+		// Write the embedded HTML content to the response
+		logsHTML, err := getHTMLFile("logs.html")
+		if err != nil {
+			c.String(http.StatusInternalServerError, err.Error())
+			return
+		}
+		_, err = c.Writer.Write(logsHTML)
+		if err != nil {
+			c.String(http.StatusInternalServerError, fmt.Sprintf("failed to write response: %v", err))
+			return
+		}
+	} else {
+		c.Header("Content-Type", "text/plain")
+		history := pm.logMonitor.GetHistory()
+		_, err := c.Writer.Write(history)
+		if err != nil {
+			c.AbortWithError(http.StatusInternalServerError, err)
+			return
+		}
 	}
 }

@@ -28,7 +48,7 @@ func (pm *ProxyManager) streamLogsHandler(c *gin.Context) {
 	notify := c.Request.Context().Done()
 	flusher, ok := c.Writer.(http.Flusher)
 	if !ok {
-		c.AbortWithError(http.StatusInternalServerError, fmt.Errorf("Streaming unsupported"))
+		c.AbortWithError(http.StatusInternalServerError, fmt.Errorf("streaming unsupported"))
 		return
 	}

@@ -38,11 +58,7 @@ func (pm *ProxyManager) streamLogsHandler(c *gin.Context) {
 	if !skipHistory {
 		history := pm.logMonitor.GetHistory()
 		if len(history) != 0 {
-			_, err := c.Writer.Write(history)
-			if err != nil {
-				c.AbortWithError(http.StatusInternalServerError, err)
-				return
-			}
+			c.Writer.Write(history)
 			flusher.Flush()
 		}
 	}
@@ -53,7 +69,7 @@ func (pm *ProxyManager) streamLogsHandler(c *gin.Context) {
 		case msg := <-ch:
 			_, err := c.Writer.Write(msg)
 			if err != nil {
-				c.AbortWithError(http.StatusInternalServerError, err)
+				// just break the loop if we can't write for some reason
 				return
 			}
 			flusher.Flush()
@@ -2,10 +2,13 @@ package proxy

 import (
 	"bytes"
+	"encoding/json"
 	"fmt"
 	"net/http"
 	"net/http/httptest"
+	"sync"
 	"testing"
+	"time"

 	"github.com/stretchr/testify/assert"
 )
@@ -31,7 +34,7 @@ func TestProxyManager_SwapProcessCorrectly(t *testing.T) {
 		assert.Equal(t, http.StatusOK, w.Code)
 		assert.Contains(t, w.Body.String(), modelName)

-		_, exists := proxy.currentProcesses["/"+modelName]
+		_, exists := proxy.currentProcesses[ProcessKeyName("", modelName)]
 		assert.True(t, exists, "expected %s key in currentProcesses", modelName)

 	}
@@ -41,21 +44,31 @@ func TestProxyManager_SwapProcessCorrectly(t *testing.T) {
 }

 func TestProxyManager_SwapMultiProcess(t *testing.T) {
+
+	model1 := "path1/model1"
+	model2 := "path2/model2"
+
+	profileModel1 := ProcessKeyName("test", model1)
+	profileModel2 := ProcessKeyName("test", model2)
+
 	config := &Config{
 		HealthCheckTimeout: 15,
 		Models: map[string]ModelConfig{
-			"model1": getTestSimpleResponderConfig("model1"),
-			"model2": getTestSimpleResponderConfig("model2"),
+			model1: getTestSimpleResponderConfig("model1"),
+			model2: getTestSimpleResponderConfig("model2"),
 		},
 		Profiles: map[string][]string{
-			"test": {"model1", "model2"},
+			"test": {model1, model2},
 		},
 	}

 	proxy := New(config)
 	defer proxy.StopProcesses()

-	for modelID, requestedModel := range map[string]string{"model1": "test/model1", "model2": "test/model2"} {
+	for modelID, requestedModel := range map[string]string{
+		"model1": profileModel1,
+		"model2": profileModel2,
+	} {
 		reqBody := fmt.Sprintf(`{"model":"%s"}`, requestedModel)
 		req := httptest.NewRequest("POST", "/v1/chat/completions", bytes.NewBufferString(reqBody))
 		w := httptest.NewRecorder()
@@ -67,10 +80,177 @@ func TestProxyManager_SwapMultiProcess(t *testing.T) {

 	// make sure there's two loaded models
 	assert.Len(t, proxy.currentProcesses, 2)
-	_, exists := proxy.currentProcesses["test/model1"]
-	assert.True(t, exists, "expected test/model1 key in currentProcesses")
-
-	_, exists = proxy.currentProcesses["test/model2"]
-	assert.True(t, exists, "expected test/model2 key in currentProcesses")
+	_, exists := proxy.currentProcesses[profileModel1]
+	assert.True(t, exists, "expected "+profileModel1+" key in currentProcesses")

+	_, exists = proxy.currentProcesses[profileModel2]
+	assert.True(t, exists, "expected "+profileModel2+" key in currentProcesses")
+}
+
+// When a request for a different model comes in ProxyManager should wait until
+// the first request is complete before swapping. Both requests should complete
+func TestProxyManager_SwapMultiProcessParallelRequests(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping slow test")
+	}
+
+	config := &Config{
+		HealthCheckTimeout: 15,
+		Models: map[string]ModelConfig{
+			"model1": getTestSimpleResponderConfig("model1"),
+			"model2": getTestSimpleResponderConfig("model2"),
+			"model3": getTestSimpleResponderConfig("model3"),
+		},
+	}
+
+	proxy := New(config)
+	defer proxy.StopProcesses()
+
+	results := map[string]string{}
+
+	var wg sync.WaitGroup
+	var mu sync.Mutex
+
+	for key := range config.Models {
+		wg.Add(1)
+		go func(key string) {
+			defer wg.Done()
+
+			reqBody := fmt.Sprintf(`{"model":"%s"}`, key)
+			req := httptest.NewRequest("POST", "/v1/chat/completions?wait=1000ms", bytes.NewBufferString(reqBody))
+			w := httptest.NewRecorder()
+
+			proxy.HandlerFunc(w, req)
+
+			if w.Code != http.StatusOK {
+				t.Errorf("Expected status OK, got %d for key %s", w.Code, key)
+			}
+
+			mu.Lock()
+
+			results[key] = w.Body.String()
+			mu.Unlock()
+		}(key)
+
+		<-time.After(time.Millisecond)
+	}
+
+	wg.Wait()
+	assert.Len(t, results, len(config.Models))
+
+	for key, result := range results {
+		assert.Equal(t, key, result)
+	}
+}
+
+func TestProxyManager_ListModelsHandler(t *testing.T) {
+	config := &Config{
+		HealthCheckTimeout: 15,
+		Models: map[string]ModelConfig{
+			"model1": getTestSimpleResponderConfig("model1"),
+			"model2": getTestSimpleResponderConfig("model2"),
+			"model3": getTestSimpleResponderConfig("model3"),
+		},
+	}
+
+	proxy := New(config)
+
+	// Create a test request
+	req := httptest.NewRequest("GET", "/v1/models", nil)
+	req.Header.Add("Origin", "i-am-the-origin")
+	w := httptest.NewRecorder()
+
+	// Call the listModelsHandler
+	proxy.HandlerFunc(w, req)
+
+	// Check the response status code
+	assert.Equal(t, http.StatusOK, w.Code)
+
+	// Check for Access-Control-Allow-Origin
+	assert.Equal(t, req.Header.Get("Origin"), w.Result().Header.Get("Access-Control-Allow-Origin"))
+
+	// Parse the JSON response
+	var response struct {
+		Data []map[string]interface{} `json:"data"`
+	}
+	if err := json.Unmarshal(w.Body.Bytes(), &response); err != nil {
+		t.Fatalf("Failed to parse JSON response: %v", err)
+	}
+
+	// Check the number of models returned
+	assert.Len(t, response.Data, 3)
+
+	// Check the details of each model
+	expectedModels := map[string]struct{}{
+		"model1": {},
+		"model2": {},
+		"model3": {},
+	}
+
+	for _, model := range response.Data {
+		modelID, ok := model["id"].(string)
+		assert.True(t, ok, "model ID should be a string")
+		_, exists := expectedModels[modelID]
+		assert.True(t, exists, "unexpected model ID: %s", modelID)
+		delete(expectedModels, modelID)
+
+		object, ok := model["object"].(string)
+		assert.True(t, ok, "object should be a string")
+		assert.Equal(t, "model", object)
+
+		created, ok := model["created"].(float64)
+		assert.True(t, ok, "created should be a number")
+		assert.Greater(t, created, float64(0)) // Assuming the timestamp is positive
+
+		ownedBy, ok := model["owned_by"].(string)
+		assert.True(t, ok, "owned_by should be a string")
+		assert.Equal(t, "llama-swap", ownedBy)
+	}
+
+	// Ensure all expected models were returned
+	assert.Empty(t, expectedModels, "not all expected models were returned")
+}
+
+func TestProxyManager_ProfileNonMember(t *testing.T) {
+
+	model1 := "path1/model1"
+	model2 := "path2/model2"
+
+	profileMemberName := ProcessKeyName("test", model1)
+	profileNonMemberName := ProcessKeyName("test", model2)
+
+	config := &Config{
+		HealthCheckTimeout: 15,
+		Models: map[string]ModelConfig{
+			model1: getTestSimpleResponderConfig("model1"),
+			model2: getTestSimpleResponderConfig("model2"),
+		},
+		Profiles: map[string][]string{
+			"test": {model1},
+		},
+	}
+
+	proxy := New(config)
+	defer proxy.StopProcesses()
+
+	// actual member of profile
+	{
+		reqBody := fmt.Sprintf(`{"model":"%s"}`, profileMemberName)
+		req := httptest.NewRequest("POST", "/v1/chat/completions", bytes.NewBufferString(reqBody))
+		w := httptest.NewRecorder()
+
+		proxy.HandlerFunc(w, req)
+		assert.Equal(t, http.StatusOK, w.Code)
+		assert.Contains(t, w.Body.String(), "model1")
+	}
+
+	// actual model, but non-member will 404
+	{
+		reqBody := fmt.Sprintf(`{"model":"%s"}`, profileNonMemberName)
+		req := httptest.NewRequest("POST", "/v1/chat/completions", bytes.NewBufferString(reqBody))
+		w := httptest.NewRecorder()
+
+		proxy.HandlerFunc(w, req)
+		assert.Equal(t, http.StatusNotFound, w.Code)
+	}
 }
Author	SHA1	Message	Date
Benson Wong	314d2f2212	remove cmd_stop configuration and functionality from PR #40 (#44 ) * remove cmd_stop functionality from #40	2025-01-31 12:42:44 -08:00
Benson Wong	fad25f3e11	Use client request context in proxy request (#43 ) Canceled or closed HTTP requests from clients will also stop the proxied HTTP requests to upstreamed servers.	2025-01-31 10:21:49 -08:00
Benson Wong	2c3e3e27f7	Support OPTIONS requests (#42 ) Add middleware that responds with permissive OPTIONS headers for all request paths.	2025-01-31 10:09:07 -08:00
Benson Wong	baeb0c4e7f	Add cmd_stop configuration to better support docker (#35 ) Add `cmd_stop` to model configuration to run a command instead of sending a SIGTERM to shutdown a process before swapping.	2025-01-30 16:59:57 -08:00
Benson Wong	2833517eef	Improve handling of process that do not handle SIGTERM (#38 ) - Process TTL goroutine did not have a return after .Stop() - Improve logging - Add test TestProcess_LowTTLValue to measure SIGTERM error rate	2025-01-20 14:39:52 -08:00
Benson Wong	abdc2bfdb3	Fix panic when requesting non-members of profiles A panic occurs when a request for an invalid profile:model pair is made. The edge case is that the profile exists and the model exists but they're not configured as a pair. This adds an additional check to make sure the profile:model pair is valid before attempting to swap the model.	2025-01-16 12:06:38 -08:00
Benson Wong	c3b834737f	Update README.md	2025-01-13 22:37:30 -08:00
Benson Wong	3c8e727b73	Update README.md	2025-01-12 19:48:35 -08:00
Benson Wong	3a1e9f81f1	support TTS /v1/audio/speech (#36 )	2025-01-12 16:27:01 -08:00
Benson Wong	72c883f36c	Update README.md	2025-01-02 09:01:51 -08:00
Benson Wong	1b04d034cf	Update README.md	2025-01-02 08:59:11 -08:00
Benson Wong	2e45f5692a	Update README.md Improve README documentation.	2025-01-01 12:51:24 -08:00
Benson Wong	c97b80bdfe	Update README.md	2025-01-01 12:25:45 -08:00
Benson Wong	ae3ef9bc39	Refactor UI (#33 ) - add html to / instead of 404 - add client side regex to /logs	2024-12-23 19:48:59 -08:00
Benson Wong	db6715bec3	update golang.org/x/net -> v0.33.0 for dependabot	2024-12-20 11:28:32 -08:00
Benson Wong	da5d9e8a6a	fix HTTP logging so true path is printed	2024-12-20 11:25:01 -08:00
Benson Wong	84b667ca7a	improve logging and error reporting for troubleshooting	2024-12-20 10:46:56 -08:00
Benson Wong	29657106fc	add more OpenAI API supported in README	2024-12-20 10:08:20 -08:00
Benson Wong	9c8860471e	support v1/rerank endpoint	2024-12-17 21:22:25 -08:00
Benson Wong	9b4e3f307e	rename proxy handler	2024-12-17 17:25:10 -08:00
Benson Wong	6fe37c3abf	support /v1/embeddings (#4 )	2024-12-17 17:25:10 -08:00
Benson Wong	7f45493a37	Update README.md	2024-12-17 14:45:41 -08:00
Benson Wong	891f6a5b5a	Add /upstream endpoint (#30 ) * remove catch-all route to upstream proxy (it was broken anyways) * add /upstream/:model_id to swap and route to upstream path * add /upstream HTML endpoint and unlisted option * add /upstream endpoint to show a list of available models * add `unlisted` configuration option to omit a model from /v1/models and /upstream lists * add favicon.ico	2024-12-17 14:37:44 -08:00
Benson Wong	7183f6b43d	fix bad logging due to wrong []byte used #28	2024-12-16 16:22:14 -08:00
Benson Wong	d89bfeb441	add .DS_Store to .gitignore	2024-12-16 12:30:31 -08:00
Benson Wong	9a0c6bed40	Improve stop exceptions (#28 ) (#29 ) Stop Process TTL goroutine when process is not ready (#28) - fix issue where the goroutine will continue even though the child process is no longer running and the Process' state is not Ready - fix issue where some logs were going to stdout instead of p.logMonitor causing them to not show up in the /logs - add units to unloading model message	2024-12-16 12:29:25 -08:00
Benson Wong	d6ca535939	tweak release tagging so it is not based on number of commits	2024-12-14 15:46:10 -08:00
Benson Wong	27302c0c02	change llama-swap to use goreleaser default ldflag values	2024-12-14 10:30:06 -08:00
Benson Wong	d4e22cceaa	Fix security vulnerability with golang.org/x/crypto - does not affect the project as llama-swap does not use the crypto libraries - good practice to keep security deps updated!	2024-12-14 10:20:22 -08:00
Benson Wong	4c94927658	Move release to Makefile out of goreleaser - less complexity - easier - goreleaser, github, pipelines: 1... mostlygeek: 0	2024-12-14 10:16:46 -08:00
Benson Wong	a955a4a5c0	create tag to release	2024-12-14 10:07:20 -08:00
Benson Wong	22d3f1a4f9	Change versioning to use git commits counts instead of semver - less work for me - more frequent releases	2024-12-14 09:53:13 -08:00
Benson Wong	e2443251ad	update readme	2024-12-09 19:14:49 -08:00
Benson Wong	5fbd53c616	delay TTL check until after all requests are complete (#25 ) - fixes #25 where requests that last longer than the TTL will cause the process to be unloaded before the next request. - new behavior, TTL waits until all requests are complete before checking timeout	2024-12-09 19:08:03 -08:00
Benson Wong	97dae50dc4	update readme	2024-12-08 21:34:16 -08:00
Benson Wong	cb978f760f	add web interface to /logs	2024-12-08 21:26:22 -08:00
Benson Wong	387f0ef6c4	use new timings data in server response in run-benchmark.sh	2024-12-03 20:48:36 -08:00
Benson Wong	18c134624d	Add Access-Control-Allow-Origin CORS header to /v1/models endpoint - match behavior of llama.cpp where the Origin in request is used - add test for listModelsHandler	2024-12-03 15:53:59 -08:00
Benson Wong	da2326bdc7	add example: optimizing code generation	2024-12-03 10:25:43 -08:00
Benson Wong	da46545630	fix profile example in README	2024-12-01 10:13:31 -08:00
Benson Wong	04b4760e7e	change profile split character to : (colon) (#21 ) - change from `/` to `:` for multiple models loaded as part of a profile - breaking change now, but allows for more compatibility with other inference engines that may have model references like `coding:Qwen/Qwen-2.5-Coder-32B`	2024-12-01 09:10:50 -08:00
Benson Wong	9fc5d5b5eb	improve cmd parsing (#22 ) Switch from using a naive strings.Fields() to shlex.Split() for parsing the model startup command into a string[]. This makes parsing much more reliable around newlines, quotes, etc.	2024-12-01 09:02:58 -08:00
Benson Wong	cf82b3c633	Improve Concurrency and Parallel Request Handling (#19 ) Rewrite the swap behaviour so that in-flight requests block process swapping until they are completed. Additionally: - add tests for parallel requests with proxy.ProxyManager and proxy.Process - improve Process startup behaviour and simplified the code - stopping of processes are sent SIGTERM and have 5 seconds to terminate, before they are killed	2024-11-30 15:24:42 -08:00
Benson Wong	e363f8f498	clean up writing with AI :b	2024-11-28 22:12:44 -08:00
Benson Wong	c9629cf3a2	add speculative decoding example	2024-11-28 22:07:22 -08:00
Benson Wong	50426935a4	.	2024-11-28 22:06:29 -08:00
Benson Wong	2fceb78e8d	Add examples	2024-11-28 22:05:41 -08:00
Ikko Eltociear Ashimine	9a81c53664	chore: update process_test.go (#17 ) nonexistant -> nonexistent	2024-11-26 10:20:16 -08:00
Benson Wong	716d37de82	Update README.md fix grammar	2024-11-25 12:35:00 -08:00