Hello, appears that @cf/meta/m2m100-1.2b is down or consistently returning 429s anyway, failing a bu
Hello, appears that @cf/meta/m2m100-1.2b is down or consistently returning 429s anyway, failing a bunch of workers I have
�this GA model is a night mare, unstable and truncate input unexpectedly
TYeah I would love a couple of other translating and embedding models in CF but everything seems to be about text and image generation at the moment
SHi
i am stuck with this error can anyone help me
Tensor error: failed to build tensor image: Tensor error: Unknown internal error: failed to decode u8
code:
app.post("/ai-process", async (c) => {
// Parse the form data from the request
const formData = await c.req.parseBody();
console.log("1", formData);
const text = formData.text || "Default text";
const image = formData.image;
console.log("2", image);
if (!image) {
return c.json({ error: "No image uploaded or invalid format" }, 400);
}
const imageArrayBuffer = await image.arrayBuffer();
const u8 = new Uint8Array(imageArrayBuffer);
// Prepare the AI request
const aiRequest = {
messages: [
{
role: "user",
content: text, // e.g., "What is the purchase price?"
},
],
image: Buffer.from(u8).toString("base64"),
};
console.log("3", aiRequest);
try {
// Call the AI API
const results = await c.env.AI.run(
"@cf/meta/llama-3.2-11b-vision-instruct",
aiRequest
);
console.log("4", results);
return c.json({
message: "Form processed successfully!",
response: results.response,
});
} catch (error) {
return c.json(
{
error:
},
500
);
}
});
i am stuck with this error can anyone help me
Tensor error: failed to build tensor image: Tensor error: Unknown internal error: failed to decode u8
code:
app.post("/ai-process", async (c) => {
// Parse the form data from the request
const formData = await c.req.parseBody();
console.log("1", formData);
const text = formData.text || "Default text";
const image = formData.image;
console.log("2", image);
if (!image) {
return c.json({ error: "No image uploaded or invalid format" }, 400);
}
const imageArrayBuffer = await image.arrayBuffer();
const u8 = new Uint8Array(imageArrayBuffer);
// Prepare the AI request
const aiRequest = {
messages: [
{
role: "user",
content: text, // e.g., "What is the purchase price?"
},
],
image: Buffer.from(u8).toString("base64"),
};
console.log("3", aiRequest);
try {
// Call the AI API
const results = await c.env.AI.run(
"@cf/meta/llama-3.2-11b-vision-instruct",
aiRequest
);
console.log("4", results);
return c.json({
message: "Form processed successfully!",
response: results.response,
});
} catch (error) {
return c.json(
{
error:
AI processing failed: ${error.message},},
500
);
}
});
Sam I crazy or is
@cf/qwen/qwen1.5-14b-chat-awq returning blank responsesSworks with other models
SSeems to work 90% of the time, sometimes it still returns blank responses
SNot too sure why or how to reproduce it consistently
SNone defined, just providing 3 inputs (system/user/system)
SI use it to run some tests to compare to an llm I'm building, so it'll send 5 queries in ~2 seconds using the rest api
SMight be some sort of output handling thing? It seems to process & generate, but just doesn't return it
FAny clue why
@hf/meta-llama/meta-llama-3-8b-instruct is returning the prompt in the response with these tags? Previously it appeared the typical behaviour was to return the actual response which in our scenario would include all text excluding <|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\nGenerate a joke from the users response.<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n. Can we expect such inconsistencies in a non-beta model? 
IThanks for reporting, we were able to reproduce and have escalated this. 
IIt's definitely not intentional.
FCheers, appreciate it man!
FWould it be safe to say that the cloudflare workersAI are safe for production?
CHey there! Thanks for bringing this to our attention. We are taking a look into this now.
CShould be fixed now
Fcheers people appreciate the punctuality
Fall is swell on my end, thanks
BAny update?
BWas told we would get a rix over a month ago, super disappointing
B
Dunfortunately I have no new information, sorry :/
BThanks are you even a cloudflare developer? 
(Sorry for accusing you or anything)
(Sorry for accusing you or anything)Bhahah the green role makes you look very official
DGreen names are not employees, we're community volunteers.
Dwith that, I still try my best to find answers, and answer questions to the best of my ability; unfortunately in this case we'd need someone from the team to update us on the status
Bit be whatit be
GHi
I want to use this model too.
Please tell me how to use the model to generate Japanese audio.
I want to use this model too.
Please tell me how to use the model to generate Japanese audio.
I^ this - we have a number of private channels to the respective teams so we can ask them questions but don't have internal access, so we're generally reliant on them to investigate things internally
0Hi I used @cf/meta/llama-3-8b-instruct which is suppose to be free but I got alert daily free allocation has been used up?
Cyea it just hates short prompts in general, when it first was added there was lots of silly false positives like "elbow" "a elbow" "female" all being blocked for nsfw, didn't recheck but I assume it hasn't changed and would still false flag those, would nice to have the option to disable
Chow do I see AI usage?
Gokay, thank you!
VWondering the same, according to cloudflare it will "soon" be visible in the dashboard -> https://developers.cloudflare.com/workers-ai/platform/pricing/
Cloudflare Docs
Workers AI is included in both the Free and Paid Workers plans and is priced based on model task, model size, and units.
JHas anyone had any luck transcribing audio and fixing the target language to something other than english? Seems like
source_lang acts more as a suggestion than a rule. Not sure if this is a Cloudflare or a Whisper thing, but figured I'd ask. Relevant snippet: 
NWhat's the best way to diagnose the issue that's causing
@hf/meta-llama/meta-llama-3-8b-instruct to intermittently return blank responses?
Gafaik the Resnet50 will create a 2048 dimensions per vector, but the vectorize seems has limitations of 1536 dimensions per vector, do you have any tips on reducing the dimensions?
SPlease do not ping community members for non-moderation reasons. Doing so will not solve your issue faster and will make people less likely to want to help you.
RBeen a while, but last time I checked the model had a limit of ~2800 tokens (input + output). Longer than that and you'd get a blank response. Most of the models give blank or truncated responses when reaching their limits, so that's the first thing I'd check.
BIs there a way to get the number of tokens consumed by each request, both for input and output token?
SIs there any zero-shot classification model available in Workers AI models, such as the one provided by Hugging Face (e.g., facebook/bart-large-mnli)?
AAny chance Claude Haiku or Sonnet will come to workers ai?
IWorkers AI models are hosted by Cloudflare (all are open-source models), they're not external providers like Claude or OpenAI, so no they won't be coming.
