On longs pages it seems to get stuck about lumos

[andrewnguonly]

This issue is definitely on my radar. At the moment, the Ollama embeddings API can only handle 1 request at a time. It's a known issue and I believe it's being prioritized (see ollama/ollama#358).

Some kind of progress indicator would be nice. Material UI has a Linear determinate progress bar component: https://mui.com/material-ui/react-progress/#linear-determinate. This should be pretty easy to implement (i.e. message passing between background script and main extension).

Separately, I'm working on a small refactor to make things a little bit faster: #52

[sublimator]

        const vectorStore = await MemoryVectorStore.fromDocuments(
            documents,
            new OllamaEmbeddings({
              baseUrl: OLLAMA_BASE_URL,
              model: OLLAMA_MODEL,
            }),
        );

I think fromDocuments/OllamaEmbeddings runs serially anyway so may need some wrapper beyond the server proxy.

[sublimator]

Probably worth investigating yourself
I really like the magic / "just works" aspect of using RAG/embeddings, but it would be nice if it was a bit faster /$somehow/

[sublimator]

I mean for cloud, it definitely would help.
I just realized that's for OpenAI / Cohere

[sublimator]

Have been pondering IPFS global shared embeddings cache
The great big CAS in the sky

It seems so wasteful for everyone to be doing this separately

[sublimator]

Assume you saw this:
https://www.youtube.com/watch?v=Ml179HQoy9o

[sublimator]

So much nicer after Nomic anyway :)

[sublimator]

I wonder if one could make a little node.js prototype server for this that simply wraps/proxies to multiple llama.cpp instances using mmap mode. Would be fun to see it purring and get an inkling of future performance.

[sublimator]

Looks like someone made an Ollama proxy server:
https://github.com/ParisNeo/ollama_proxy_server

I've never tinkered much with the ollama memory settings ( so job well done Ollama team)
Can it use mmap?

[sublimator]

Seems it can use mmap. At least for the /api/generate there is a use_mmap option.
You can start multiple instances of ollama by setting OLLAMA_HOST environment variable.
e.g. OLLAMA_HOST=0.0.0.0:11435 ollama serve

[sublimator]

Don't know if there's "anything on the table" regardless

[sublimator]

Apparently the embeddings don't use the entire weights, so maybe there's a way. I'm very fuzzy on how those are created.

I patched the proxy server to allow CORS, but I am not having much luck, and suddenly it's giving me crazy Yoda responses :)

Maybe there's a way to start a pool of ollama instances used just for the embedding, which don't ever load the full model weights and allocate too many buffers (ofc mmap doesn't solve everything)

[sublimator]

[GIN] 2024/01/31 - 10:15:31 | 200 |  1.460126417s |       127.0.0.1 | POST     "/api/embeddings"
127.0.0.1 - - [31/Jan/2024 10:15:31] "POST /api/embeddings HTTP/1.1" 200 -
127.0.0.1 - - [31/Jan/2024 10:15:31] "POST /api/embeddings HTTP/1.1" - -
[GIN] 2024/01/31 - 10:15:31 | 200 |  1.462187292s |       127.0.0.1 | POST     "/api/embeddings"
127.0.0.1 - - [31/Jan/2024 10:15:31] "POST /api/embeddings HTTP/1.1" 200 -
127.0.0.1 - - [31/Jan/2024 10:15:31] "POST /api/embeddings HTTP/1.1" - -
[GIN] 2024/01/31 - 10:15:31 | 200 |  1.418453666s |       127.0.0.1 | POST     "/api/embeddings"
127.0.0.1 - - [31/Jan/2024 10:15:31] "POST /api/embeddings HTTP/1.1" 200 -
127.0.0.1 - - [31/Jan/2024 10:15:31] "POST /api/embeddings HTTP/1.1" - -
[GIN] 2024/01/31 - 10:15:31 | 200 |  1.458442375s |       127.0.0.1 | POST     "/api/embeddings"
127.0.0.1 - - [31/Jan/2024 10:15:31] "POST /api/embeddings HTTP/1.1" 200 -
127.0.0.1 - - [31/Jan/2024 10:15:31] "POST /api/embeddings HTTP/1.1" - -
[GIN] 2024/01/31 - 10:15:32 | 200 |  1.416166584s |       127.0.0.1 | POST     "/api/embeddings"
127.0.0.1 - - [31/Jan/2024 10:15:32] "POST /api/embeddings HTTP/1.1" 200 -
127.0.0.1 - - [31/Jan/2024 10:15:32] "POST /api/embeddings HTTP/1.1" - -
[GIN] 2024/01/31 - 10:15:32 | 200 |  1.461118458s |       127.0.0.1 | POST     "/api/embeddings"
127.0.0.1 - - [31/Jan/2024 10:15:32] "POST /api/embeddings HTTP/1.1" 200 -
127.0.0.1 - - [31/Jan/2024 10:15:32] "POST /api/embeddings HTTP/1.1" - -
[GIN] 2024/01/31 - 10:15:32 | 200 |  1.460722667s |       127.0.0.1 | POST     "/api/embeddings"
127.0.0.1 - - [31/Jan/2024 10:15:32] "POST /api/embeddings HTTP/1.1" 200 -
127.0.0.1 - - [31/Jan/2024 10:15:32] "POST /api/embeddings HTTP/1.1" - -
[GIN] 2024/01/31 - 10:15:32 | 200 |   1.50568075s |       127.0.0.1 | POST     "/api/embeddings"

Hacked the ollama langchain code to do things in parallel and sadly not really getting any speed up. Maybe due to the threads settings for the servers. Maybe it's already pretty efficient in using available cores?

I was getting around 250ms per embed, which went up to around 1.5s per embed.

Maybe there is a tiny speedup, shrug, but seemingly not a low hanging fruit substantial one.

[andrewnguonly]

@sublimator, thanks for testing out all of that stuff! For reference, how large is the content in your testing (i.e. how much text is on the page)? And about how many embeds before 250ms goes to 1.5s? In my testing, I didn't observe an increase in latency; it seems to be constant throughout the entire sequence of embeds.

If there's a comparable Wikipedia article, we can both work off of that for testing.

[sublimator]

@andrewnguonly

If there's a comparable Wikipedia article,
how large is the content in your testing

I was kind of using random pages

it seems to be constant throughout the entire sequence of embeds.

Did you use parallel processing though? By default the OllamaEmbedding class does the requests sequentially. Let me dig it back up.

[sublimator]

I hacked the OllamaEmbeddings class (just the compiled code in node_modules):

    async _embed(strings) {
        console.log('hack is working!!')
        const embeddings = [];
        for await (const prompt of strings) {
            const embedding = this.caller.call(() => this._request(prompt));
            embeddings.push(embedding);
        }
        return await Promise.all(embeddings)
    }
    async embedDocuments(documents) {
        return this._embed(documents);
    }

[sublimator]

Maybe we should start a branch if want to look seriously, but anyway, here's some more of the artifacts of my investigations before:

The ini file I used

[DefaultServer]
url = http://localhost:11442
queue_size = 1

[SecondaryServer]
url = http://localhost:11435
queue_size = 1

[SecondaryServer1]
url = http://localhost:11436
queue_size = 2

[SecondaryServer2]
url = http://localhost:11437
queue_size = 3

[SecondaryServer3]
url = http://localhost:11438
queue_size = 4

[SecondaryServer4]
url = http://localhost:11439
queue_size = 5

[SecondaryServer5]
url = http://localhost:11440
queue_size = 6

[SecondaryServer6]
url = http://localhost:11441
queue_size = 7

proxy server hacks (I might not have been using the right python version, but all I wanted to change was the CORS stuff)

diff --git a/ollama_proxy_server/main.py b/ollama_proxy_server/main.py
index 87944d0..5f53a8e 100644
--- a/ollama_proxy_server/main.py
+++ b/ollama_proxy_server/main.py
@@ -47,9 +47,9 @@ def main():
     parser.add_argument('--log_path', default="access_log.txt", help='Path to the access log file')
     parser.add_argument('--users_list', default="authorized_users.txt", help='Path to the config file')
     parser.add_argument('--port', type=int, default=8000, help='Port number for the server')
-    parser.add_argument('-d', '--deactivate_security', action='store_true', const=True, default=False, help='Deactivates security')
+    parser.add_argument('-d', '--deactivate_security', action='store_true', default=False, help='Deactivates security')
     args = parser.parse_args()
-    servers = get_config(args.config)  
+    servers = get_config(args.config)
     authorized_users = get_authorized_users(args.users_list)
     deactivate_security = args.deactivate_security
     print("Ollama Proxy server")
@@ -58,22 +58,25 @@ def main():
     class RequestHandler(BaseHTTPRequestHandler):
         def add_access_log_entry(self, event, user, ip_address, access, server, nb_queued_requests_on_server, error=""):
             log_file_path = Path(args.log_path)
-        
+
             if not log_file_path.exists():
                 with open(log_file_path, mode='w', newline='') as csvfile:
                     fieldnames = ['time_stamp', 'event', 'user_name', 'ip_address', 'access', 'server', 'nb_queued_requests_on_server', 'error']
                     writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
                     writer.writeheader()
-        
+
             with open(log_file_path, mode='a', newline='') as csvfile:
                 fieldnames = ['time_stamp', 'event', 'user_name', 'ip_address', 'access', 'server', 'nb_queued_requests_on_server', 'error']
                 writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
                 row = {'time_stamp': str(datetime.datetime.now()), 'event':event, 'user_name': user, 'ip_address': ip_address, 'access': access, 'server': server, 'nb_queued_requests_on_server': nb_queued_requests_on_server, 'error': error}
                 writer.writerow(row)
-                
+
         def _send_response(self, response):
             self.send_response(response.status_code)
             self.send_header('Content-type', response.headers['content-type'])
+            self.send_header('Access-Control-Allow-Origin', '*')
+            self.send_header('Access-Control-Allow-Methods', 'GET, POST, OPTIONS')
+            self.send_header('Access-Control-Allow-Headers', 'Authorization, Content-Type')
             self.end_headers()
             self.wfile.write(response.content)
 
@@ -81,6 +84,13 @@ def main():
             self.log_request()
             self.proxy()
 
+        def do_OPTIONS(self):
+            self.send_response(200, "ok")
+            self.send_header('Access-Control-Allow-Origin', '*')
+            self.send_header('Access-Control-Allow-Methods', 'GET, POST, OPTIONS')
+            self.send_header('Access-Control-Allow-Headers', 'Authorization, Content-Type')
+            self.end_headers()
+
         def do_POST(self):
             self.log_request()
             self.proxy()
@@ -93,7 +103,7 @@ def main():
                     return False
                 token = auth_header.split(' ')[1]
                 user, key = token.split(':')
-                
+
                 # Check if the user and key are in the list of authorized users
                 if authorized_users.get(user) == key:
                     self.user = user
@@ -112,11 +122,11 @@ def main():
                 if not auth_header or not auth_header.startswith('Bearer '):
                     self.add_access_log_entry(event='rejected', user="unknown", ip_address=client_ip, access="Denied", server="None", nb_queued_requests_on_server=-1, error="Authentication failed")
                 else:
-                    token = auth_header.split(' ')[1]                
+                    token = auth_header.split(' ')[1]
                     self.add_access_log_entry(event='rejected', user=token, ip_address=client_ip, access="Denied", server="None", nb_queued_requests_on_server=-1, error="Authentication failed")
                 self.send_response(403)
                 self.end_headers()
-                return            
+                return
             url = urlparse(self.path)
             path = url.path
             get_params = parse_qs(url.query) or {}
@@ -141,16 +151,17 @@ def main():
             if path == '/api/generate':
                 que = min_queued_server[1]['queue']
                 client_ip, client_port = self.client_address
+                self.user = 'unknown'
                 self.add_access_log_entry(event="gen_request", user=self.user, ip_address=client_ip, access="Authorized", server=min_queued_server[0], nb_queued_requests_on_server=que.qsize())
                 que.put_nowait(1)
                 try:
                     response = requests.request(self.command, min_queued_server[1]['url'] + path, params=get_params, data=post_params)
                     self._send_response(response)
                 except Exception as ex:
-                    self.add_access_log_entry(event="gen_error",user=self.user, ip_address=client_ip, access="Authorized", server=min_queued_server[0], nb_queued_requests_on_server=que.qsize(),error=ex)                    
+                    self.add_access_log_entry(event="gen_error",user=self.user, ip_address=client_ip, access="Authorized", server=min_queued_server[0], nb_queued_requests_on_server=que.qsize(),error=ex)
                 finally:
                     que.get_nowait()
-                    self.add_access_log_entry(event="gen_done",user=self.user, ip_address=client_ip, access="Authorized", server=min_queued_server[0], nb_queued_requests_on_server=que.qsize())                    
+                    self.add_access_log_entry(event="gen_done",user=self.user, ip_address=client_ip, access="Authorized", server=min_queued_server[0], nb_queued_requests_on_server=que.qsize())
             else:
                 # For other endpoints, just mirror the request.
                 response = requests.request(self.command, min_queued_server[1]['url'] + path, params=get_params, data=post_params)

#!/bin/bash

# Start the Default Server
OLLAMA_ORIGINS=* OLLAMA_HOST=0.0.0.0:11442 ollama serve &

# Start Secondary Servers
OLLAMA_ORIGINS=* OLLAMA_HOST=0.0.0.0:11435 ollama serve &
OLLAMA_ORIGINS=* OLLAMA_HOST=0.0.0.0:11436 ollama serve &
OLLAMA_ORIGINS=* OLLAMA_HOST=0.0.0.0:11437 ollama serve &
OLLAMA_ORIGINS=* OLLAMA_HOST=0.0.0.0:11438 ollama serve &
OLLAMA_ORIGINS=* OLLAMA_HOST=0.0.0.0:11439 ollama serve &
OLLAMA_ORIGINS=* OLLAMA_HOST=0.0.0.0:11440 ollama serve &
OLLAMA_ORIGINS=* OLLAMA_HOST=0.0.0.0:11441 ollama serve &


ollama_proxy_server --config config.ini --port 11434 -d

[sublimator]

how large is the content in your testing

Enough that there was a lot of embeddings requests anyway

This was one of the pages:
https://news.ycombinator.com/item?id=39197619

But I suspect it's grown since I was testing

[sublimator]

I would try hacking a separate pool of servers just for the embeddings, with the proxy running on a non-default port. I'm still not sure WHEN the full model weights are loaded, but from the logging it could very well be lazily when the generate api call is hit

[andrewnguonly]

I hacked the OllamaEmbeddings class (just the compiled code in node_modules):

    async _embed(strings) {
        console.log('hack is working!!')
        const embeddings = [];
        for await (const prompt of strings) {
            const embedding = this.caller.call(() => this._request(prompt));
            embeddings.push(embedding);
        }
        return await Promise.all(embeddings)
    }
    async embedDocuments(documents) {
        return this._embed(documents);
    }

I feel like this change (or something similar) will be accepted in LangChainJS. See:

1. langchain-ai/langchainjs#2140
2. langchain-ai/langchainjs#2294

[sublimator]

In any case, it didn't seem to help much in the big picture.

Maybe you can tweak the threading settings for each ollama instance or something

[sublimator]

I mean that might actually work somehow - cause the shared embeddings would just share public data. And then you just need your little local LLM for processing the private queries. I guess it would get complicated by needing to trust that content -> embeddings for a given model, but I suppose they could be signed. Shrug. Want FAST responses, and to waste less energy $SOMEHOW

[sublimator]

https://ollama.com/library/nomic-embed-text

[andrewnguonly]

The latest version of Ollama seems to be much more performant when loading/unlocking models. Having separate inference and embedding models will feel smoother. I'll close this issue after a couple more updates are implemented (e.g. batch embedding creation).