Invalid index error when generating model

[mirkov]

I get invalid index when generating the model

This is on a MacBook M2 with SBCL 2.5.0

Here are the steps that lead to the error (Please see my comments at bottom of post):

CL-USER> (ql:quickload :lla )
(:LLA)
CL-USER> (ql:quickload :llama )
(:LLAMA)
CL-USER> (in-package :llama )
#<PACKAGE "LLAMA">
LLAMA> (init #P"stories15M.bin" #P"tokenizer.bin" 3200 )
; No values
LLAMA> (generate *model* *tokenizer*)
; Debugger entered on #<SB-INT:INVALID-ARRAY-INDEX-ERROR expected-type: (INTEGER 0 (3200)) datum: 9038>
[1] LLAMA>
; Evaluation aborted on #<SB-INT:INVALID-ARRAY-INDEX-ERROR expected-type: (INTEGER 0 (3200)) datum: 9038>

The error is

Invalid index 9038 for (SIMPLE-VECTOR 3200), should be a non-negative integer below 3200.
   [Condition of type SB-INT:INVALID-ARRAY-INDEX-ERROR]

Restarts:
 0: [RETRY] Retry SLY mREPL evaluation request.
 1: [*ABORT] Return to SLY's top level.
 2: [ABORT] abort thread (#<THREAD tid=5891 "sly-channel-1-mrepl-remote-1" RUNNING {7004FB05E3}>)

Backtrace:
 0: ((SB-VM::OPTIMIZED-DATA-VECTOR-REF T) #("<unk>" " ..)
      Locals:
        SB-INT:INDEX = 9038
        VECTOR = #("<unk>" "\n<s>\n" "\n</s>\n" "<0x00>" "<0x01>" "<0x02>" "<0x03>" "<0x04>" "<0x05>" "<0x06>" "<0x07>" "<0x08>" "<0x09>" "<0x0A>" "<0x0B>" "<0x0C>" "<0x0D>" "<0x0E>" "<0x0F>" "<0x10>" "<0x11>" ..)
 1: (GENERATE #<TRANSFORMER  {700610F523}> #<TOKENIZER  {700610F563}> :TOPP NIL :TEMPERATURE 0.9 :STEPS 256 :PROMPT NIL)
      Locals:
        #:.DEFAULTING-TEMP. = 256
        #:.DEFAULTING-TEMP.#1 = NIL
        MODEL = #<TRANSFORMER  {700610F523}>
        NEXT-TOKEN = 9038
        POSITION = 0
        PROMPT-TOKENS = #()
        TEMPERATURE = 0.9
        TOKENIZER = #<TOKENIZER  {700610F563}>
        TOPP = NIL
 2: (SB-INT:SIMPLE-EVAL-IN-LEXENV (GENERATE *MODEL* *TOKENIZER*) #<NULL-LEXENV>)
 3: (EVAL (GENERATE *MODEL* *TOKENIZER*))
 4: ((LAMBDA NIL :IN SLYNK-MREPL::MREPL-EVAL-1))
 --more--

• The same error appears with different size tokens. I tried with 10000.
• My installation of lla and binary-format is flaky so maybe I should not even expect llama to function correctly. lla tests drop me to low level debugger. binary-format tests fail.
• May be this is a MacBook M2 platform related issue. I will try llama on the x86 platform next week.

I submitted issues for lla and binary-format.