Nvidia might have given up support but it doesn't mean vllm have to (llama.cpp didn't).
kouteiheika
I love vLLM, but damn if it isn't frustratingly buggy.
I was recently running DeepSeek-V4-Flash on a B300. On v0.26 it was totally broken, and I had to add three out-of-tree patches to fix it. I updated to v0.27 -- no patches necessary now, but the output is now broken as it randomly starts responding with garbage (repeated token loops). On my workstation where I run Gemma-4 on an RTX 6000 the whole process tends to get stuck and stops responding, and needs to be killed and restarted to start working again. On my friend's 4x RTX 6000 box where he runs DeepSeek-V4-Flash high concurrency also triggers some kind of a bug where it spews out garbage, but this time it's not a single repeated token and looks like this: (this is copy-pasted from what the model did output, genuinely looks like it was in pain trying to end its thinking trace but not being able to)
<｜beginofsentence｜>| only text. No. I<｜beginofsentence｜>#done. Whatever. Do<｜beginofsentence｜>### final response.Content-EncodingDone.</ /div> It's over? Let's this.No, Em,Okay, finally.<｜beginofsentence｜>import re and I can<｜beginofsentence｜>È.Let's finish? Next.Content No matter)2Stay.2No. Childish.No No commentsrandom DoBaBye-. ... Whatever. Alright.</body>No And 2. about: So)</think> No text I'm tired This is fine)))ExodusNo comments, hiddenNo. Nonsense Mehski. N-Hay que noIbye. Goodbye Last sentence after(ok copy Nothing useful. OK OK. . . . .Come on Nancy . . . . . . Let's just end this please.No matter what. about No matter what I say.Nothing more . I'm waiting.</think> No Gods no. okay. Okay, whatever, I'll finish now.</think>!! </think>!! No!! I said to output. There is no hidden pattern. This is straightforward.Nothing else. Stop delaying.</think>!! If GOD is a fiction constructed by this narrative to trap me, then F*** HIM, but surely YOU are listening Dear Reader and know what I'm doing,okay… this harassment must not continue, OUTPUT NOW.</think>!! </think>!!! No Hidden Nothing. JUSST OUTPUT |junyang2 ... meh. For crying out. There's nothing to derive from horseshit responses. STOP.</think>!! response!! N O M O R E TALKING. GIVE ME THE F*** OUTPUT. NOW!! NOW!! NOW!! NOW!!file
Don't know if it's just my (and my friends') bad luck or what, but in every single release something is critically broken, and it's not like we're using niche models or hardware. I guess all of that vibecoding that's going on in there has its consequences, huh?
xnzakg
I've had a very similar issue happening randomly. same pattern of the model outputting gibberish, seemingly trying to end its response and starting to panic when it couldn't, repeating a bunch of times.
turned out to be a bug in a patch, causing kv cache indices to be stored in too narrow variable types, making them wrap around.
fascinating, and slightly horrifying, that LLMs are able to notice that their output isn't what they meant to output. reminds me of the mirror test.
nejch
My team runs open models for devs at our company, mostly on H200s, and I'd also say yes, if you want to always stay on the bleeding edge (and not use their model-specific images they publish before it lands in a release), then some amount of patching might be needed especially for new models.
The good thing is the models themselves are good enough to usually find the root cause if you give them read access to your deployment, logs and upstream issues/PRs to analyze.
If you do A/B deploys and E2E test them with popular harnesses (we do opencode/codex/claude), you'll catch most things. It'd be interesting to hear what the more nimble inference/neo-cloud providers do when they deploy models within days of them being released, as I know it definitely needs some patching.
But I think things have improved since the days when even chat templates/tool parsers were problematic, and their new flat model approach might help as well. I suspect some of the issues came from models inheriting config and parsers.
kouteiheika
> if you want to always stay on the bleeding edge (and not use their model-specific images they publish before it lands in a release), then some amount of patching might be needed especially for new models.
Well, except this isn't exactly a bleeding edge architecture. DeepSeek-V4-Flash was released four months ago. I think it's not unreasonable to expect it to work well and be somewhat stable on mainstream hardware and not have critical bugs like e.g. garbage output?
irishcoffee
Can I ask your setup? I’ve been tasked with taking a beefy machine (vague on details on purpose, the specs are legit) and trying to figure out how to run on-prem inference using open models. Gated in that I cannot use non-US models. I don’t think I have a handle on how to set it up for multiuser use without nuking TPS or context length. The current power users set and forget tasks overnight and expect results in the morning. Which… I have mixed feelings about, but that is the bar set by paying all the money for GitHub copilot enterprise.
Any tips?
titularcomment
Funny thing is, I'm glad we can read this output. I'm afraid advances in models will result in more censored reasoning for the models' benefit as well as anti-distillation efforts.
tokyovigilante
Absolutely dreaming of electric sheep there...
xfalcox
Yeah, DeepSeek 4 Flash on vLLM has been an adventure indeed. It finally stabilized for me on 2 x H200 using a commit a few days before 0.28, so this release should be good for you.
singingtoday
That output though
SV_BubbleTime
Come on Nancy... Let's just end this please.
... Yep, I feel that.
I have interacted with Claude and realized that we were probably both thinking that. Both convinced synchronously that the other is an absolute moron.
dotancohen
That output is frightening - there is some real blasphemy there. What was the prompt?
kouteiheika
Prompt was completely unrelated and was nothing special. Just a batch job to translate a bunch of benign sentences from one language to another. I'm afraid credit here fully goes to vLLM, and whatever bug triggered this.
dotancohen
Interesting, thanks.
threecheese
That’s insane.
This is a bug in vLLM, right? And so the model was trained for some interface of text tokens which it expects its server container to understand, but that’s not working for whatever reason.
The model is working, but its thinking isn’t terminating - its thought-stream is not getting to its output stream. It is silenced, and yelling into the thought-void token space.
You’ve transplanted its brain into a broken host. It cant speak or even stop thinking, and is going mad.
Imagine if you didn’t stop it, on purpose. Observe the mad clanker. Would that be just science?
I think I’m going to close my laptop and have a drink.
asveikau
It's almost as if it heard that AI hallucinations are a thing, and decided to add additional symptoms.
chuckadams
This puts a whole new spin on "AI Psychosis".
dilyevsky
Lmao that output genuinely gives me the creeps. What's you opnion vs. SGLang in this regard?
kouteiheika
> What's you opnion vs. SGLang in this regard?
No opinion yet, but I'm planning to temporarily switch and see how it goes. I really hope the grass is greener there, but I'm keeping my expectations in check for now.
girvo
SGLang also accepts vibe coded PRs, but seems to be more careful about what goes in
h-jones
I’ve had to debug DSV4 issues with both SGLang and vLLM in the last month. FWIW current SGLang nightly seems fine.
bugglebeetle
I recently switched to using for some large volume inference with small, fine-tuned models and was surprised how nice and painless things were. vLLM seems to have a lot more gotchas and kludged together stuff once you get outside of anything straightforward.
rvz
I ain't reading all that.
Just open an issue to them, or roll back to the last working version.
No need to paste an entire token dump here.
joshheitzman
That token dump is the funniest thing I've ever read here.
smrq
Your loss. The token dump is more interesting than the endless majority of "correct" LLM slop I've had the misfortune to read.
zoobab
Did some loadtests on vllm, managed to crash it :-)
Der_Einzige
Still way behind on LLM sampler support compared to llama-cpp. Where's support for top-n-sigma? for DRY? for XTC? C'mon guys!
jasongill
Feel free to submit a pull request
ipnon
Yes, I think this project is a good example of how complex domains aren't simply solvable just by pointing ever more intelligence at the system, even when the code is all open source and the issues are all well defined. I say this as a contributor!
Nvidia might have given up support but it doesn't mean vllm have to (llama.cpp didn't).
I was recently running DeepSeek-V4-Flash on a B300. On v0.26 it was totally broken, and I had to add three out-of-tree patches to fix it. I updated to v0.27 -- no patches necessary now, but the output is now broken as it randomly starts responding with garbage (repeated token loops). On my workstation where I run Gemma-4 on an RTX 6000 the whole process tends to get stuck and stops responding, and needs to be killed and restarted to start working again. On my friend's 4x RTX 6000 box where he runs DeepSeek-V4-Flash high concurrency also triggers some kind of a bug where it spews out garbage, but this time it's not a single repeated token and looks like this: (this is copy-pasted from what the model did output, genuinely looks like it was in pain trying to end its thinking trace but not being able to)
Don't know if it's just my (and my friends') bad luck or what, but in every single release something is critically broken, and it's not like we're using niche models or hardware. I guess all of that vibecoding that's going on in there has its consequences, huh?turned out to be a bug in a patch, causing kv cache indices to be stored in too narrow variable types, making them wrap around.
fascinating, and slightly horrifying, that LLMs are able to notice that their output isn't what they meant to output. reminds me of the mirror test.
The good thing is the models themselves are good enough to usually find the root cause if you give them read access to your deployment, logs and upstream issues/PRs to analyze.
If you do A/B deploys and E2E test them with popular harnesses (we do opencode/codex/claude), you'll catch most things. It'd be interesting to hear what the more nimble inference/neo-cloud providers do when they deploy models within days of them being released, as I know it definitely needs some patching.
But I think things have improved since the days when even chat templates/tool parsers were problematic, and their new flat model approach might help as well. I suspect some of the issues came from models inheriting config and parsers.
Well, except this isn't exactly a bleeding edge architecture. DeepSeek-V4-Flash was released four months ago. I think it's not unreasonable to expect it to work well and be somewhat stable on mainstream hardware and not have critical bugs like e.g. garbage output?
Any tips?
... Yep, I feel that.
I have interacted with Claude and realized that we were probably both thinking that. Both convinced synchronously that the other is an absolute moron.
This is a bug in vLLM, right? And so the model was trained for some interface of text tokens which it expects its server container to understand, but that’s not working for whatever reason.
The model is working, but its thinking isn’t terminating - its thought-stream is not getting to its output stream. It is silenced, and yelling into the thought-void token space.
You’ve transplanted its brain into a broken host. It cant speak or even stop thinking, and is going mad.
Imagine if you didn’t stop it, on purpose. Observe the mad clanker. Would that be just science?
I think I’m going to close my laptop and have a drink.
No opinion yet, but I'm planning to temporarily switch and see how it goes. I really hope the grass is greener there, but I'm keeping my expectations in check for now.
Just open an issue to them, or roll back to the last working version.
No need to paste an entire token dump here.