+
�_
jz � � � R t ^ RIHt ^ RIt^ RIt^ RIt^ RIHt ^ RIt^ RI H
t
. ^N^N^*NRNRNRNRNR NR
NRNRNR
NRNRNRN^dN^�NRNRNRNRNRNRNRNRNRNRNRNRNRNRNRNR NR!NR"NR#NR$NR%NR&NR'NR(NR)NR*NR+NR,NR-NR.NR/NR0NR1NR2NR3NR4NR5NR6NR7NR8NR9NR:NR;NR<NR=NR>NR?NR@NRANRBNRCNRDNRENRFNRGNRHNRINRJNRKNRLNRMNRNNRONRPNRQNRRNRSNRTNRUNRVNRWNRXNRYNRZNR[NR\NR]NR^NR_NR`NRaNRbNRcNRdNReNRfNRgNRhNRiNRjNRkNRlNRmNRnNRoNRpNRqNRrNRsNRtNRuNRvNRwNRxNRyNRzNR{NR|NR}NR~NRNR�NR�NR�NR�NR�NR�NtR� R� ltR� R� lt
R� R� ltR� R� lt]P ! 4 R� R� l4 tR� R� ltR� R� lt]R�8X d
]! 4 R# R# )�z�Python Transformers Qwen3 timing and CPU/MPS parity matrix.
Uses the same token grid as the Rust Qwen3 matrix example. The output is
CSV-compatible and intentionally mirrors the Rust columns where possible.
)�annotationsN)�Iterable)�AutoModelForCausalLMi: i�
i)# i.j i� i�� i90 � i i i @ i � i, i� i� iX i� i i� i� i� i� i� i� ip iX i@ i(# i'