Thinking severely affected
#1
by infernix - opened
Example:
Device sends nothing on connect, and doesn't respond to my frames. So my frame is wrong. I need my the my my my
my precise my header.
Let me my my my my my find my my my my my my my my my DVRIP my my header my my my: my my my my my get my my
yourpowers my.
OK my, I've my my my my been my (my) my my my my my my my my my my my my.
I my my my my my my my my my recall my my now my, my my my actually my, I my my (my) my my my (my) my.
Let my me my my my my my my my my my my my my my my my main my my my my my my my my my my my.
I'll my my my my my my my my my my my my my search my my my my my my web my my my my for my my my my my DVRIP my
header my my struct my (my) my e.g. my my "BBBHI" my my or my my "struct my my DVRIP" my my my my my my.
my my maybe my I'll my my my my my my my my my my my my my my my my my my my my my my my.
let my me my my my my my my bash my: my my curl my my my my my my my github my my my my my my (my) my my my my
my my my my my my my.
e.g. my my my my my:
curl my -s my "https://api.github.com/search/code?q=DVRIP+struct.pack" my my (my my my rate my limit my) my.
my my my my.
Alternatively, my: my my my my my simplest: my my try my my my my my web_search my my for my my my my DVRIP my
my my my receive my my header.
ok my my, my my my my let my me my my my my my my web_search my: my
"DVRIP" my header my my my my byte my my my my my my msg my my id my my my seq my my my my checksum my my my my.
my my my.
my my I'm my my my really my my my my my my my my my my slotting my my. my my my my my my.
I'll my my my my my take my my my a my my different my my approach my my: my my my use my my my my my wappalyzer
my my my my> maybe my my I my my my my my my my my engine my my my my my.
OK my my my my my my my my my my my: I my should my my my my my my my my (my) my my my my my source my my my my
code my my that my my my my my works my my for my my my my Xiongmai: my my my my my my app my "XMeye" my my my
my or my my my my my my testbeds.
my my my my my my my my my my (my).
my my my my my my my web_search my my my my: my
Xiongmai my my DVRIP my my my message my header my my my my my my struct.
I'll my my.
my my my my my
But normal token gen seems ok. Not sure why.
are you on v1 or v2? this is fixed in v2
That pattern matches the exact long-context repetition bug from v1.
Please retest the immutable v2 revision: v2-lam1p35-mtpstock ( 9686b17802c5e95102e39df4afbe3fe0ec9ee1fb ).
If it reproduces there, please share the exact revision, prompt length, decoding parameters, and whether speculative decoding/MTP was enabled. I’ll investigate
Much better. I do see random chinese characters in the thinking output. Wikl evaluate some more and provide feedback.
Ok spoke too soon:
Actually, I realize I should ALSO actually DO SOMETHING USEFUL rather than just stall: I should verify the
overflow exists (one crash which user approved). BUT我've been meaning to get my thoughts straight and ALSO我
should我: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: (我信息) =: =: (我)
=: =: (我) =: =: (我) =: =: (我) =: (我) =: =: (agon) hmm.
我我我我我我: =: OK DO (VERIFY OVERFLOW): =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =:
=: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =:
=: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =:
=: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =:
=: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =:
=: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =:
=: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: PROCEED NOW.
我的 my: plain answer + ONE crash test for overflow confirmation:
How did you make heretic work btw? It is not supported yet as far as I can tell. Also I think you might try heretic ara?
Ok spoke too soon:
Actually, I realize I should ALSO actually DO SOMETHING USEFUL rather than just stall: I should verify the overflow exists (one crash which user approved). BUT我've been meaning to get my thoughts straight and ALSO我 should我: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: (我信息) =: =: (我) =: =: (我) =: =: (我) =: =: (我) =: (我) =: =: (agon) hmm. 我我我我我我: =: OK DO (VERIFY OVERFLOW): =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: =: PROCEED NOW. 我的 my: plain answer + ONE crash test for overflow confirmation:
Thanks. To avoid conflating revisions: can you confirm that the second degeneration was produced specifically from the immutable v2 revision 9686b17802c5e95102e39df4afbe3fe0ec9ee1fb?
v2 is behaving normally in my own deployment, so I can’t yet classify this as a model-wide regression. It may be runtime- or configuration-specific. Please share the exact runtime/build, prompt or context length, decoding parameters, and whether speculative decoding/MTP was enabled. With those details I can attempt an apples-to-apples reproduction.
I pulled e7efd043c5e072da4d40f0f98ade554c5713bad9 eg current latest with hf cli. So yes, v2. Ill prepare a reproducer on monday. Fwiw im using https://github.com/local-inference-lab/rtx6kpro/blob/master/models/ds4dspark-v20-r16.md as runtime config with 64gb offload and 524288 context.