ajb d dl mZ d dlmZmZ d dlZd dlZd dlZd dlm Z de
de e
fdZ dej
j fd Zd
e
deee e f fd Zd Zd
ZddddefdZed ZdefdZdefdZdefdZdS ) )contextmanager)AnycastN)Timerfilenamereturnc d}d}d g }t | 5 }| | }t | D ]\ }}|dk r | | }|dk r(|d | } ||dz
d fd| d D }
| d |
d d d n # 1 swxY w Y |S )
Nzzr c > g | ]}|t d S N)len).0xpfxs `/home/longshao/multi-rider-rag/.venv/lib/python3.11/site-packages/torch/utils/jit/log_extract.py
z extract_ir.. s' GGGaQs3xxyy\GGG T)keepends )openreadsplit enumeratefind
splitlinesappendjoin) r BEGINENDgraphsf
split_strsi split_strend_locslinesr s @r
extract_irr) sQ E
C
C
F
h *1VVXX^^E**
%j 1 1 * *LAyAvv nnS))G"}}(7(#A QU#..004CGGGG1<<<+F+FGGGE MM"''%.. ) ) ) ) * * * * * * * * * * * * * * * * Ms CC==DDinp_typec Z | }| }| }| }|t d |t d |t d |t d t j |||| S )NzFmake_tensor_from_type: 'size' is None (inp_type.sizes() returned None)zJmake_tensor_from_type: 'stride' is None (inp_type.strides() returned None)zImake_tensor_from_type: 'device' is None (inp_type.device() returned None)zGmake_tensor_from_type: 'dtype' is None (inp_type.dtype() returned None))sizestridedevicedtype)sizesstridesr. r/ AssertionErrortorch
empty_strided)r* r, r- r. r/ s r make_tensor_from_typer5 s >> D
F
__
F NN E | efff
~ ijjj
~ hiii } fggg De T T TTr irc t j | d }| g }| D ]}t | t j j r)| t j
dd ]t | t j j r)| t j dd t | t j j
rUt t j j
| }| t | =t | t j j r.| t j dd dk t# d| t j d| }t j |j ||fS ) NT)parse_tensor_constantsg?d r r z,A default value is not implemented for type forward)r3 _Cparse_irmakeMultiOutputIntoTupleinputs
isinstancetype FloatTyper randomuniformIntTyperandint
TensorTyper r5 BoolTypeNotImplementedError_create_function_from_graph!_jit_pass_erase_shape_informationgraph)r6 rK r> inp
tensorTypefuncs r load_graph_and_inputsrO - s H b > >E ""$$$
F||~~ c c chhjj%("4 5 5
c MM&.S11 2 2 2 2
EH$4
5
5 c MM&.C00 1 1 1 1
EH$7
8
8 c eh 1388::>>J MM/
;; < < < <
EH$5
6
6 c MM&.A..!3 4 4 4 4%&aUXU]U]U_U_&a&abb b 8 / / 5 A AD H..tz::: &>r c f t d| |d }| }|j dz S )Nz fn(*inputs))fnr> )stmtglobals )r blocked_autorangemedian)rQ r> test_runsttimess r time_cudarZ B s:
=*G*GHHHA
! !E <$ r c t j }t | D ]} | | t j }||z
|z dz S )NrT )time perf_counterrange)rQ r> rW r' _es r time_cpura G sV A
9
F A
EY %%r
) warmup_runsrW c ( t | \ }}t | D ]} || d }|D ].}t |t j r|j j dk } n/|t d |rt ||| nt ||| }|S )NcpuzNo tensor found in inputs)
rO r^ r? r3 Tensorr. r@ r2 ra rZ ) r6 r> rd rW rK r_ is_cpuinputouts r run_testrk N s $R((HE1
;
v
F eU\ * * \&%/F E ~ 899906
_(5&)
,
,
,IeVU^<_<_C Jr c / K t j d } d V t j | d S # t j | w xY w)NF)r3 r; _get_graph_executor_optimize)argskwargs old_optimizes r no_fuserrq ^ sa 888??L<
--l;;;;;--l;;;;s A !A)c r t 5 t | | cd d d S # 1 swxY w Y d S r
)rq rk r6 r> s r run_baseline_no_fusionrt f s} $ $F##$ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $s ,00c |rdgndg}t j | }t j d 5 t | | cd d d t j | S # 1 swxY w Y t j | d S # t j | w xY w)N)DYNAMICrb )STATICrb fuser1)r3 jitset_fusion_strategyfuserrk )r6 r> dynamicstrat old_strats r run_nncr k s 1%,B !!>2BI11%88
Y__X
&
& ( (B'' ( ( ( ( ( ( ( %%i0000 ( ( ( ( ( ( ( ( ( %%i00000 %%i0000s0 AB1 B B1 BB1 B B1 1!Cc t j d 5 t | | cd d d S # 1 swxY w Y d S )Nfuser2)r3 ry r{ rk rs s r run_nvfuserr t s " " $ $F##$ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $s =AA)
contextlibr typingr r rB r3 r\ torch.utils.benchmarkr strlistr) r; rF r5 tuplerO rZ ra floatrk rq rt r r r r r s % % % % % %
' ' ' ' ' ' c *
UEH$7
U
U
U
U c eCcN&; *
& & & )+b U < < <$% $ $ $ $
1E 1 1 1 1$u $ $ $ $ $ $r