From 8f089af0d6c88481d3d6a0561f7826df609e26eb Mon Sep 17 00:00:00 2001
From: Emil Kosz <ekosz@int.pl>
Date: Sat, 21 Mar 2026 17:52:23 +0100
Subject: [PATCH] Major speed improvement, split the project into several files

---
 __pycache__/chat.cpython-313.pyc        | Bin 0 -> 3090 bytes
 __pycache__/preferences.cpython-313.pyc | Bin 0 -> 1931 bytes
 __pycache__/prompt.cpython-313.pyc      | Bin 0 -> 4656 bytes
 __pycache__/util.cpython-313.pyc        | Bin 0 -> 534 bytes
 chat.py                                 |  43 ++++++
 main.py                                 | 127 +++++++++++++++++
 preferences.py                          |  24 ++++
 prompt.py                               |  65 +++++++++
 pyllama.py                              | 174 ------------------------
 util.py                                 |   5 +
 10 files changed, 264 insertions(+), 174 deletions(-)
 create mode 100644 __pycache__/chat.cpython-313.pyc
 create mode 100644 __pycache__/preferences.cpython-313.pyc
 create mode 100644 __pycache__/prompt.cpython-313.pyc
 create mode 100644 __pycache__/util.cpython-313.pyc
 create mode 100644 chat.py
 create mode 100644 main.py
 create mode 100644 preferences.py
 create mode 100644 prompt.py
 delete mode 100644 pyllama.py
 create mode 100644 util.py

diff --git a/__pycache__/chat.cpython-313.pyc b/__pycache__/chat.cpython-313.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..7fdb7b89d30c28a4fcaf156a07cd2054ff819258
GIT binary patch
literal 3090
zcmey&%ge>Uz`)RUW?!ZdD+9x05C?`?p^VQ#3=9lY8G;#t8NC^bKs1v#qZd;VlNWOl
zvlmMdivmcD(VMl1Rf0i@A(&Z;A($nGIfw_UABi$$hH#45(-}2cUxGM(n#{LY+<h{W
zZZW%;WZz<TPsvUOiNi1xl=0bxfq@~8VJbrmLlI*zLlILjV-a&Gg9&3WlN3W1M-dCi
z-Xhju7Ab}zwqTB6Rw;%o#v=A$HYtW8j$rm8&QJzh#v(2i22IXe9L^buC1IJV<uBP7
z7#NZnp;mw>kf-?>7#KcVF)}dNF-~QO28n`k2y-YyEDHmJ5<@7X0YfO03!@n$$m^jD
zIWR{uM1u?js}AA-Qz(QIgDI3B%8&!H3Cs&+2m(g}m;)z*c;QSq1IctgxFDPnBmif^
z8KzLbDKOx$MHsFW&Io2i(IpC(fHSb0h0vAGpvmMHqRD-WwYan(wdfW{e0*kJW=VYf
zE&k-vqWHwj_@dO}g8aPV)FKuJ28LVgKAFWOPWctL*mE+AOOo;{ZV45qmc$pQ=A<T<
zWaj6^=jNxR-r^00@qIyD7C(PK*IT?Gg=zW8rNxO!IjOfeLEPkw%$yWWmRl^vsX1x4
zSQ3lUi*K=ImxF1}`1s_U#Ny)k_+pUN3JMAhx198gO7rw{Q&Tb%^;2^*bMympGP5(|
zU2`*Y67>U$^0QL2ODaK(^rFPv{PM)S%vAk=N}rs>+(iB4jKmVXg32OpkUs=r=}Hgg
zVRi-vhT<#+28JIE40kwqJJ~OBNX$^Vz@fQ7?GlIP19{~RmMa`GPsF9BS5K;*T-(9Y
z!*hd!r=O>j=Q@Y_MGo~f%-1z6FKSp`*RZ>&VRwbY9xkhNkwa;L$rTQb2OK>8e4Ts~
zN~YILthvge@);CO7-<^hkk2jPj2y}k1Wgqnl`srTnR!g1OhM3W3X=oTp-e&W6bs@*
zFick{QxM1oh#;JV$?9@3Fd(Onm5hFxEVua6Q%mB@5>v8Di%W7-^GYC*SC*JlntF>9
zBv)DhN<6$EW<g?lYJ72KRjMX)kq83=LlG!d-eOKn$tni9RY5_aNCFfC0>#Ni`8hf9
zCHe75`6VU!xwj<Yae^)-1d1C21_p)(h8uiB*ZGt#@+n>BQ|Vy7!NJ|n+sQj2d3x@|
z+{yVXpmA_n!}_9_^;Hg=A}RDz1C*3N?)dD4Cxn9;gPANDLm6`5p&!Z+1TQ(@fySW1
zpvmm#rpb7VtvEF=CACPC=@xflN=kfgYH@L5dTJ4<T-9U*2SYK)D5%%@ic1oUO5i!`
zmN>%WC=y~IZ>vDOEw6lCUjL%J{)W;E^7<F#ZNaJW2EXuie&vh&$`?44AzlS}4dy;j
z0SDrL4r2t>3W(wnn^%+=f*FGuLzy(7x)Yfg7=oGn8H1UFSuC0Jm_wPdC8|)Q0tl(d
z!&;z*G6%DoLIY6&sj$XYq#}ynJeE)vYz1s6Ga{LWG6eBKosC2V2_kWz%up7Du4quI
z0*eQM(kYmQAe2DGFem~LLSPm`e<-sGgC?6_NIF=Y;e$Pt!N8!&e2WWGB7#dgmhz&+
zf?I4Ai8-0+c`rd>qsap)O<8>slTveTae#`#lGKWlTP&p|nK`%kK>XaqqV&wX_{5x?
zTU^jWHz>cn2viFcfqZ+5EwP{=H7}(I6vekVGE;I=;}cU-ij)`_7&KYIY3>#a$j~BH
zP<CU@0b9jflwV#9DjgI+0Gtzx)Ijng@HB{&fI+#ZxD1>C?+6G_=bXg3AY=o}F}Iz;
z7XpGW3WRj9-r(Sc6lqsDWbSbAT<4I!#34Pyb%Eg(4%H`og8g}2d6)U*5E(V~3Wp55
zV3fYdAw9$S3WvfI4xWDAF5U?#*Tu9iifLcr(7C}c(!m2s($FLW4rO@mJp;?VU}01O
zRHCMXq7PLV%)?q<GJu**U}+=*TXKh$lA(;c49Mjq!K9D1-GE3+=?q0OplD+Qml%-J
z6%t?I(iO}E$8(Vhs6+)vF%vY33lfV;ia}unD_8|V5eTcb<3ZB5q);LkMGh4C#opj3
zzQN6Zom>7Q2&v!T5qbcMPMM1wGS@lOE^??XD7mg~eNo-|y1MH{b=S-4?k7|(s{38x
z@V|o^c~?1f!11QZ1ojL#t!gseVgNNVIO5~;5_41I<8N`rgIWruIUqJqe0*VPVh&V>
zKR!M!v#7WvCo?ZKFF!sWTwjAzb&)#A^B@No*@BX+XncHeNn%N6a(rS*Nl|7}X-R5v
zeEcn5XuBafH76%NUX!s%98`>gieCpvdIwd1MIZuHKo%>4s{A&FuS}e*s-MJ|SZzOX
zFtMt9;oxEw{Vu`Es{4U~l~ot)Bd{J#j$2IWnYVb;Gxds63-XIIOY(~<A#nproJF9#
zRRpT4Zt(?!E4b8@@XWlF{Bp4Fl?)J*z%?L;O>TZlX-=wLkrx941IU5J$_xw)AD9^#
m8E-Ri-e=Ib$e{6<f%^-KDx+>k@pXR1i~NdT8B`e|&H(^LH(^x(

literal 0
HcmV?d00001

diff --git a/__pycache__/preferences.cpython-313.pyc b/__pycache__/preferences.cpython-313.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..12bc8ea7e5ec31a8155fddbcab69aa33d699680a
GIT binary patch
literal 1931
zcmey&%ge>Uz`(HT%)U$(W(J1GAPx+(LK&aE7#J9)G6XXOGkP-=foLXgMlYr!CIt|i
z(VMx5S%N`{A(&Z;A($nGDTo8A4~a5mhH#2l(-}2cUxGM(nvA!Y-Al5QK^z!{Dr5Mp
z!@$50$1s&4hM|Ztn4yR%n6ZdCm?f0KgfW;&iXn?5i?N6$n7N2Gl);v<h)sn-ll7Lc
zZ+=Q@PC!v=T53^hUUF(!W@`CMHU<WUWG0yL3=9mM3=9nX3=9mP&oMGE*fE0ah=xmt
zu!k_jvM?|xF@!Q1FoZI@Fq$zcFoZIOFyz3^hz1!6)`88VN(@Ny;GhL7g%d$Ma3-7)
z%oxmM$&|+u%?A?-5`eJ~v=T!oOE9x3)Ex@Q_6Q+V!B|QR!7Rb7mW+9<p)5h7Fi`{@
z%4*68;h~r$1`$IhQG6tUECXRHF$A*(vj=koa|Ux+vIMgSa|iQSGUu^{vIR*)R3MYV
zyx=(G1H~c2FF4%DhHzsDLy#=8IS_Uz3qnOYgC@UU5h!vrnQn3C7iFgAl_ZvA=I7mF
zFD^(-&df`H2}+n%T<!s(3O<RIsYS(_jJH^G5|dJMUV>6Y5h!S18Zj_1yaWY}Cd)0>
zvc#Oy)LX1M`Q@oaw^&OH3Q~)1@fDY(7Q|=fB^RaUrskD|Xfog8N=(TrEiM6ZZ?UCh
zre~HE-(t^5ttd#$OR3@o8={+>k(igBnvx7M8sc~c1{MYe24PTIbYcLd#i<M-j6ulq
zP<5dUL9l!Q6#-GEAcBE`0Tu!(3@aJ^G?{Mk=I6zy7nH_BEW9O<o>~$Qc0zn&aeQW8
zi6+x6mg0iUyjv{AsX1xIAWtha-15{fD$Uc+O-;#6)KAUL%+U|X$;{4-cg@YrNz@N0
z%Fjy8E~x}D(u)#v^UD+SGE?;fDt&Sia})Io;90I%ub}c4hfPjmZc<93-7P&VIZ`h^
zJ~J<~BtBlxCMQ2RF{jv052hZJOp7HM7#MyuFl;EfZti)}-1CIWb+7n~Uh$W`60evi
zeFphilm8ZLaVaP&IACTKfzo9WD3ES3JLOm0;`T?1evUBLpb$@IN1t0Pr6rj;xA=-v
zOX71Ai_$am;uCXnZn64+lE5vt#Dap<yp&s9!3CLlPNgL!`FXdv91#)Ao}8bTmzrF1
zixZ?jIU_SCMU$fll*GZd-(pEDN-w^}mR$~}nUnG>intjV7;bTZA_E-YobmC=If=!^
z@$to=l$f9Z1`S0r3=9lKiXgQz_yUZbfq|iTGXn#|&jyA&9K4<EmpCM5s9fOCT%dM|
zL-U5LRtL)!4yh-6g8k`T=}VNC>o3&5u4Hvl$?AaPMLxF<<~tlb*EyswaY)Z_ULbgd
zL-~n-@N~{eoC|m^3g~vQ-r(Tw=j!B|V0wi^@`<q6bh}A*Yl7Fut%|!LBr?NsLC_Ay
zkIYPZTn~hVXM}uYWfD>D;P}A7z$tj0L*gQb#0>c>T1i(pl3}J;T;Y&;z`@hc*U2}b
z_$r5VkvszfgC-+5q>98rZ1!6m@$q?yxvBB-x47ctbMsS5b3kmK`1r!o#2ly$e|&se
zW>IlTPG(+eUVeN$B$!0w<BLIgJ2^hFq@*Y_sk9`uI6nRsFEo6UQ*(0S<24zJ3_&3V
z%1{oFFa}j<MIZuHN)+>gQd%3scP2qrnXe3jtTJFng5@>YZZV~2-r`Nq)GJCY$S=+;
z$uFt|r>`PaP&}~*<QL?Zr51q|uVjE|1V!U54jYJ1?22p|7#Kiy7t1j)FnnNUWMsU}
ez<HNJ_AvwZ7ZxE#nHkFSwP$L7We{S7SPTGNo3}y$

literal 0
HcmV?d00001

diff --git a/__pycache__/prompt.cpython-313.pyc b/__pycache__/prompt.cpython-313.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d6c766faeb8694d6cab8359ab01b05c1ca2015ab
GIT binary patch
literal 4656
zcmey&%ge>Uz`$U6W?yDI2Lr=n5C?`?p^VRN3=9lY8G;#t8NC^bKs1v#qZd;VlNWOl
zvjT$zgAzk9vlK%xOAJ#G2UHIdWy%cU6tSi=YO=lrar`uyZZW%;WZz<TPsvUOiNG)u
zl<_%<fq@~8VJbrmLlI*zLlILjV-a&Gg9&3WlN3W1M-fX5E67e$sPZDVU=}HcBKBZb
zDTX4BU^Xd+BF<oTDTX4hU=AsUBJN;LDTX4RU@j?!BHm!`BEC=tTgD=O6$VY7TU-G}
z`MCupVVSArFIgBE7?K&GW`Zbo1_lOx1_p-D6-*2ac8pUQqCuiy9KssP5X-{Apu`Z$
zXuuH4<icpisK5{nG6F0g#0I8N2qlJK#$YBeX0~JrWypc6M$!k45EN~o0A>khwPeg=
z4rLDFMU?~bf>}(Vu2x_OWex(l7$l5@gV|722qH;AnV}3=BM)vak_&~PT97Cu23S<&
zv4pY&i6TisnZfL#3_)U04iaSwjdBGB>@EnBL{bc8h9dbF$z4)VX(URCA($hW(~>ci
zA%_np9Lf+R17jm-tZwECWeAc(sDZIyZVqM05rBz>G6X5Z*a$k587bVctHA0SgsyZ3
zO>V!JJPZsBFF`)gWVywilbTi%pPN{eo|$)xt0cbw#^x!?OwWMJ@FwM#l;r0^B||hB
zZ?R{jRum-WrM#TRz`#(&ot;{#TTql*T%4NnOCUX8H!roKM7OvyHz_|SGg*`I7H4L1
zeqMZDVs7d!_T-$*<m}WGO`%(?#ia$QMYlNO<1_OzOXA~;K-v2iM@VV~I1k-oD*)%G
zTYSOEMfo{7sVU)^c`5njw`7V_OX3TP@(U8v6H8L#^Abx+ixP9<Gg3irz9j~d%}uO`
zPtMOPNzE&PiF1O)lQS}NQf>*kmZj#EIOpe;fXyxPPOZGfo}8bTmzrF1i!VPfK0CEC
z9_pW4e2FP3V6z}<Z?QR*mXze@-QtC~E<OpuWzWxxFHX%%xy7ASk{6$mn3s~1T6Bvo
zv7jI|FGZ8>7E5tzPTDP&#G>@#TWs0oU>asbG1z~%_(0ZzTn)8}Gd?~!C$YFVKE4=K
z0w{n$!z~y6qS8G5+|-oJME%s<%pCoIoXqUZc-P#_oJ9SAqWrAX?2<|lBfTgwH@`eF
zFEdp?pwcHNF*i{k;$*#o$|6v<x+RP#u=HS&04l1Bzc4T`{Aggf!@=9heu+b3hROvF
z%>`<gI5eM#Nlwq1l(V8}N74-gqXSAGnHk017&};cxWBM5NXlH7(7Gg{wZY_~ghdDM
z4Gz)k97Y#8j5e5D;jjSnbS`q}tO&lsVF>2QU*wQqz_OfsA@>yyohOnq*Cq8YN$Ovg
zH0<EJ!69~?L+&Dn+>F{4DcAKJFX}m7;c&Xa!PC#($$gze=^}^H0=p|5Mo(lF=10wn
zx?tpTQP#DC<qC(?4Qcu7(z+L=b+1dCUX(W7ka<AlxXeMB%hG-zX^96ABPPUN;ZTMe
zHzDM@sQN`w^{X5jpF!CTrPu})DWGKe*$iBBg)#(zl%U}tkP0+0SRtmuu#(YFlc`9Y
zfq~%`S9)qmd{SvzT58cs#v)Y)28Lo#si;t-4U%Pp$lMY{iWZ0vs5mMHMOg#G1AgKD
z>dxxx{HhoERWI|aU*J$L(nYUqK<)wI&sHS5I+ziqxM1?T#h}SlWCU^@2PprPq*jy^
z8KSwHH903WvFMfnlC!}gpyIArgn@yffdLxS*EwV^a>&dGy2_z|>1a^O13P*LiH^n*
zDIoK3L<$ZapgJ0dj$lSu<0FU@O$h@7yv9;t&}8y+(`1E)7$~+uA$E%o#40XHEGmi5
zEJ-c8#RKA}=A}S*ti>fonFZ*<S_Fzalq?F;1Tsz$6#mYTaD}A?<%|5v3tX0aFZ8~y
zVs}x+?y`!*4HXRt-{zu<&1Dt41FWB!83o)K@4NdPEInNQftiuhpAj<}Kn)2{Kzue}
z1hv&d8G=Ab6wS}D)D5m`L3spCJ_t+TBd4uU25gNQY(X2$gh-J%baA6u&cFa_0D$T_
zGzo$xVQcO%hcX1gt7e3ONEJ7za0W}HGib8-xmAf46y+zU78mOlm*f|~tHB~+1_lOA
z&RbmRMTtrAY5B>e#gIaQ7nHGzQ}c>5OESw+i$Jk^3oM4HG`UhTi=p*X5vXZUBn!%s
zocVe2$%%Q%sW~f|z=cVX8YpFeqC){%q_AY<=cL@?N6MKX0dT>S14%#}JpKHg{8u<6
z;aODiB8TDv)hiqtH#m5%b4Xp}kh;#Hagjsgf~Lb24o4(G)r%ae3sSDDnO#&fyTW0P
zt>l$ED0f*ppo8TqhXkIK<^oG;;7~**u$8T7DG5~rn1?O>p`=-OoPm{tiD8lYLW)2Q
za%frur>`PM1_p*A2M_^DT18GE7O3<<q@*GxkeDla;$bVz1Em&}`UoPV15#Q9Pe9;0
z8JuiXu5hTs6B4NYzQUn^lwu&s$odM04U!-c$p}_~!;=wm@d8Tk;Cf>Uq~2gc^sItF
za^T2C5LOWk42dQT452JY>Oh4aLJ63Kt$@;HWMI$$%Q7$|GBPlvGlnu?OE!q4md=#U
zsLAZ-7NW^o1hV87m$z$Vd{AmhX;I!S9^d?w%(TqZqL9jh)LUG^9-eL?@xG41-rxw)
zWCdsHTTJnZx7e~%E6WmdZZX9t-C`{+Ni0b%1_c4AMF`31MWD*|mLyU}M{UNMfMN_p
zG%(!Y7Mv11J#J#$Wp3pg{326|uZyc+6j#44Zgf%H=sLgA4RM(pJOcebojyH&H-yD+
z@QYw(T~L((t|UNyn#z#QP^8bmz_5~`2t5&jeR4|()m@-=vJI#&4DtiGaX_pMMWNVI
zYLN@5I21v(1X42EgVcwE2v8$QlNpleL5(qRL8~bM>VAMaCh_sNxZ>k;^HWN5Qsd)q
z@x;d$mL}#vW%%Rc(=v;SOL8*vQuFfT;~~vw4p1K?wWy>LT-_I$gKPm++C`vIjv`MG
z7ZfYtAb>bgG(H{_d6~)ai6tdPnMtK3sm1Z}w|Joq{p8e~ocMT6#v(0HY=Oeq0o*by
z0+m)^0u<cE?x5Cr8^c#7VOGr#S`2J_ADI|gH9j*lu<?BWu|BaevMPRJV`BAZ{KUb^
z8p!yGj}^ogU}F{iB*4hZ`<;!EmG=V!BP%aB+`%SmvfW}z&%DK(o~c)qT999yS(0B=
z3GR_W(k-)7eg#+|Bnm*G33l5p4jV|I*%hTQFff1|T`bMO!0>^Yk&*E>1Lu7Pr7tWt
TjJy+sr;AS%|H@#)2(c6ZD)4yn

literal 0
HcmV?d00001

diff --git a/__pycache__/util.cpython-313.pyc b/__pycache__/util.cpython-313.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..b20692b6ea3a888aaee72fdec9f5fa3248863bd2
GIT binary patch
literal 534
zcmey&%ge>Uz`(HM^uEj)3=9m9K^z!ngfc#ZluTs^W{_eiVn}DyWJ+d&%77?l1_lNY
z{@laBz%Z2|lpzSF4<rl4p^T<bQ3a3+kb0;Lj6zev3KK`rXe!tdGB6ghigX4|Mn6rK
zTLQ(YCGok5Md_J&@x>*HMJ2cRkVH!I3sCr}c`3I9k>r!|OG@%{HJNU)m1m};r<UAe
z%q<3aUP0lOlYUWYo_=m>N@k*dYHntZen3uUc4oY5Ze~uRen3%vR%&)hC5Vw;l$e`e
zo|u=Jsvl74larX6s9#!=nWI-wiDGqPPEHXhph4kOY{kI9(7<qmh3h(t^hFlw8ID(2
z<iI@1i!72e6tA$zVByJJWRaN>c$G!|Gblj(G#SC>tz;<TW?*0_;$&c8xW!?Uo1ape
clWJGQ%fP?@whCnD2WCb_#?MSFj0#}20A*2gN&o-=

literal 0
HcmV?d00001

diff --git a/chat.py b/chat.py
new file mode 100644
index 0000000..d92e4ce
--- /dev/null
+++ b/chat.py
@@ -0,0 +1,43 @@
+import gi
+from gi.repository import GLib, Gtk, Gdk
+
+import util
+
+class ChatView(Gtk.ScrolledWindow):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        self.cur_ai_response = None
+        
+        self.listbox = Gtk.ListBox()
+        self.listbox.set_selection_mode(Gtk.SelectionMode.NONE)
+        self.listbox.set_focusable(True)
+        
+        self.set_child(self.listbox)
+
+    def scroll_to_bottom(self, *args):
+        adj = self.get_vadjustment()
+        adj.set_value(adj.get_upper() - adj.get_page_size())
+
+    def start_response(self):
+        self.cur_ai_response = self.add_message(sender=False)
+        return self.cur_ai_response
+        
+    def add_message(self, text="", sender=True):
+        label = Gtk.Label(selectable=True, wrap=True, xalign=1.0 if sender else 0.0)
+        label.set_text(text)
+        util.set_margin_all(label, 8)
+        
+        row = Gtk.ListBoxRow()
+        row.set_child(label)
+        self.listbox.append(row)
+        
+        GLib.idle_add(self.scroll_to_bottom)
+        return label
+
+    def add_response_part(self, part):
+        if self.cur_ai_response is None:
+            return
+
+        self.cur_ai_response.set_label(self.cur_ai_response.get_label() + part)
+        GLib.idle_add(self.scroll_to_bottom)
diff --git a/main.py b/main.py
new file mode 100644
index 0000000..a8c242d
--- /dev/null
+++ b/main.py
@@ -0,0 +1,127 @@
+import gi, ollama, asyncio, threading
+gi.require_version("Gtk", "4.0")
+from gi.repository import GLib, Gtk, Gdk
+
+from prompt import PromptView
+from chat import ChatView
+from preferences import ModelPreferenceView
+import util
+        
+class PyLlamaWindow(Gtk.ApplicationWindow):
+    def __init__(self, application=None, title=None):
+        super().__init__(application=application, title=title)	
+        
+        self.set_default_size(800, 600)
+        
+        self.model_store = Gtk.StringList()
+        self.model_dropdown = Gtk.DropDown(tooltip_text="Select model")
+        self.model_dropdown.set_model(self.model_store)
+
+        model_preferences = Gtk.Button(icon_name="view-more-symbolic",
+                                       tooltip_text="Model preferences")
+        model_preferences.connect("clicked", self.on_preferences_clicked)
+
+        self.model_prefs = ModelPreferenceView(position=Gtk.PositionType.BOTTOM,
+                                               on_gpu_change=self.on_gpu_changed)
+        self.model_prefs.set_parent(model_preferences)
+        
+        header = Gtk.HeaderBar(show_title_buttons=True)
+        header.pack_start(self.model_dropdown)
+        header.pack_start(model_preferences)
+        
+        self.set_titlebar(header)
+
+        vbox = Gtk.Box(orientation=Gtk.Orientation.VERTICAL, spacing=6)
+        util.set_margin_all(vbox, 8)
+
+        self.chat = ChatView(vexpand=True)
+        vbox.append(self.chat)
+        
+        self.prompt = PromptView(spacing=6)
+        self.prompt.on_send = self.on_send
+        self.prompt.on_cancel = self.on_cancel
+        vbox.append(self.prompt)
+
+        self.set_child(vbox)
+        self.populate_models()
+
+    def populate_models(self):
+        try:
+            models = ollama.list()
+            names = [m.model for m in models["models"]]
+        except Exception as e:
+            print ("Failed to fetch models: ", e)
+            names = []
+
+        self.model_store.splice(0, len(self.model_store), names)
+
+    def on_gpu_changed(self, value):
+        self.get_application().num_gpu = value
+        
+    def on_preferences_clicked(self, button):
+        self.model_prefs.popup()
+        
+    def on_send(self, *args):
+        text = self.prompt.get_text()
+        if not text:
+            return
+
+        self.prompt.clear()
+        self.chat.add_message(text, sender=True)
+        self.chat.start_response()
+
+        app = self.get_application()
+        model = self.model_dropdown.get_selected_item().get_string()
+
+        async def task():
+            await app.send_prompt(model, text)
+
+        if app.current_async_task and not app.current_async_task.done():
+            app.current_async_task.cancel()
+            
+        app.current_async_task = asyncio.run_coroutine_threadsafe(task(), app.async_loop)
+
+    def on_cancel(self, *args):
+        app = self.get_application()
+        app.stop_generation()
+
+class PyLlama(Gtk.Application):
+    def __init__(self):
+        super().__init__(application_id="eur.cfpi-fpsi.pyllama")
+        GLib.set_application_name("PyLlama")
+
+        self.num_gpu = 15
+
+        self.client = ollama.AsyncClient()        
+        self.async_loop = asyncio.new_event_loop()
+        self.current_async_task = None
+        threading.Thread(target=self._run_asyncio_loop, daemon=True).start()
+        
+    def do_activate(self):
+        self.window = PyLlamaWindow(application=self, title="PyLlama")
+        self.window.present()
+
+    def _run_asyncio_loop(self):
+        asyncio.set_event_loop(self.async_loop)
+        self.async_loop.run_forever()
+        return True
+
+    def stop_generation(self):
+        if self.current_async_task and not self.current_async_task.done():
+            self.current_async_task.cancel()
+            
+    async def send_prompt(self, model, text=""):
+        messages = [{ "role": "user", "content": text }]
+        GLib.idle_add(self.window.prompt.hold, True)
+        
+        try:
+            async for part in await self.client.chat(model=model, messages=messages, stream=True, options={ "num_gpu": self.num_gpu }):
+                GLib.idle_add(self.window.chat.add_response_part, part.message.content)
+        except asyncio.CancelledError:
+            raise
+        finally:
+            GLib.idle_add(self.window.prompt.unhold, True)            
+            
+if __name__ == "__main__":
+    app = PyLlama()
+    app.run()
diff --git a/preferences.py b/preferences.py
new file mode 100644
index 0000000..fcb4fb4
--- /dev/null
+++ b/preferences.py
@@ -0,0 +1,24 @@
+import gi
+from gi.repository import Gtk
+
+import util
+
+class ModelPreferenceView(Gtk.Popover):
+    def __init__(self, on_gpu_change=None, *args, **kwargs):
+        super().__init__(*args, **kwargs)    
+        self.on_gpu_change = on_gpu_change
+        
+        box = Gtk.Box(orientation=Gtk.Orientation.VERTICAL, spacing=6)
+        util.set_margin_all(box, 8)
+
+        label = Gtk.Label(label="GPU Layers")
+        box.append(label)
+        
+        gpu_spin = Gtk.SpinButton(adjustment=Gtk.Adjustment(
+            value=15, lower=0, upper=50, step_increment=1,
+        ), digits=0, hexpand=True)
+
+        gpu_spin.connect("value-changed", lambda spin: self.on_gpu_change(spin.get_value_as_int()))
+        box.append(gpu_spin)        
+        
+        self.set_child(box)
diff --git a/prompt.py b/prompt.py
new file mode 100644
index 0000000..2c6a8cd
--- /dev/null
+++ b/prompt.py
@@ -0,0 +1,65 @@
+import gi
+from gi.repository import Gtk, Gdk
+
+class PromptView(Gtk.Box):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        self.prompt = Gtk.TextView(left_margin=12, top_margin=14,
+                                   right_margin=12, bottom_margin=14)
+        
+        prompt_scroll = Gtk.ScrolledWindow(hexpand=True)
+        prompt_scroll.set_propagate_natural_height(True)
+        prompt_scroll.set_max_content_height(150)
+        prompt_scroll.set_child(self.prompt)
+        
+        key_controller = Gtk.EventControllerKey()
+        key_controller.connect("key-pressed", self.on_key_pressed)
+        self.prompt.add_controller(key_controller)
+        
+        self.prompt_button = Gtk.Button(icon_name="go-next-symbolic")
+        self.btn_handler = self.prompt_button.connect("clicked", self.on_send)
+
+        self.append (prompt_scroll)
+        self.append (self.prompt_button)
+
+    @property
+    def buffer(self):
+        return self.prompt.get_buffer()
+
+    def clear(self):
+        self.buffer.set_text("")
+
+    def get_text(self):
+        return self.buffer.get_text(self.buffer.get_start_iter(),
+                                    self.buffer.get_end_iter(), False).strip()
+        
+    def hold(self, *args):
+        self.grab_focus()
+        self.prompt.set_sensitive(False)
+        self.prompt_button.set_icon_name("process-stop-symbolic")
+        self.prompt_button.disconnect(self.btn_handler)
+        self.btn_handler = self.prompt_button.connect("clicked", self.on_cancel)
+
+    def unhold(self, *args):
+        self.prompt.set_sensitive(True)
+        self.prompt.grab_focus()
+        self.prompt_button.set_icon_name("go-next-symbolic")
+        self.prompt_button.disconnect(self.btn_handler)
+        self.btn_handler = self.prompt_button.connect("clicked", self.on_send)
+        
+        
+    def on_key_pressed(self, _a, keyval, _b, state):
+        if keyval == Gdk.KEY_Return:
+            if state & Gdk.ModifierType.SHIFT_MASK:
+                return False
+            self.on_send()
+            return True
+        return False
+
+    def on_send(self, *args):
+        pass
+
+    def on_cancel(self, *args):
+        pass
+
diff --git a/pyllama.py b/pyllama.py
deleted file mode 100644
index 833cbe9..0000000
--- a/pyllama.py
+++ /dev/null
@@ -1,174 +0,0 @@
-import gi, ollama, asyncio
-gi.require_version("Gtk", "4.0")
-from gi.repository import GLib, Gtk
-
-def set_margin_all(widget, m):
-    widget.set_margin_start(m)
-    widget.set_margin_top(m)
-    widget.set_margin_end(m)
-    widget.set_margin_bottom(m)
-
-class PyLlamaWindow(Gtk.ApplicationWindow):
-    def __init__(self, application=None, title=None):
-        super().__init__(application=application, title=title)	
-        self.set_default_size(800, 600)
-        self.cur_ai_response = None
-        self.gpu_spin = None
-        
-        self.model_store = Gtk.StringList()
-        self.model_dropdown = Gtk.DropDown()
-        self.model_dropdown.set_model(self.model_store)
-        self.model_dropdown.set_tooltip_text("Select model")
-
-        model_preferences = Gtk.Button()
-        icon = Gtk.Image.new_from_icon_name("view-more-symbolic")
-        model_preferences.set_child(icon)
-        model_preferences.set_tooltip_text("Preferences")
-        model_preferences.connect("clicked", self.on_preferences_clicked)
-
-        header = Gtk.HeaderBar()
-        header.set_show_title_buttons(True)
-        header.pack_start(self.model_dropdown)
-        header.pack_start(model_preferences)
-        
-        self.set_titlebar(header)
-        
-        vbox = Gtk.Box(orientation=Gtk.Orientation.VERTICAL, spacing=6)
-        set_margin_all(vbox, 8)
-        
-        self.listbox = Gtk.ListBox()
-        self.listbox.set_selection_mode(Gtk.SelectionMode.NONE)
-        self.listbox.set_focusable(True)
-        
-        scrolled = Gtk.ScrolledWindow()
-        scrolled.set_vexpand(True)
-        scrolled.set_child(self.listbox)
-        vbox.append(scrolled)
-
-        # self.prompt = Gtk.TextView()
-        # self.prompt.set_hexpand(True)
-        
-        self.entry = Gtk.Entry()
-        self.entry.set_placeholder_text("Type message and press Enter...")
-        self.entry.set_hexpand(True)
-        self.entry.set_icon_from_icon_name(Gtk.EntryIconPosition.SECONDARY, "go-next-symbolic")
-        self.entry.connect("activate", self.send_message)
-        self.entry.connect("icon_press", self.send_message)
-
-        vbox.append(self.entry)
-        self.set_child(vbox)
-
-        self.current_model = None
-        self.populate_models()
-
-    def hold_prompt(self, *_):
-        self.entry.set_placeholder_text("Wait for response...")
-        self.listbox.grab_focus()
-        self.entry.set_sensitive(False)
-
-    def unhold_prompt(self, *_):
-        self.entry.set_placeholder_text("Type message and press Enter...")
-        self.entry.set_sensitive(True)
-        self.entry.grab_focus()
-        
-    def on_preferences_clicked(self, button):
-        box = Gtk.Box(orientation=Gtk.Orientation.VERTICAL, spacing=6)
-        set_margin_all(box, 8)
-
-        label = Gtk.Label(label="GPU Layers")
-        label.set_xalign(0.0)
-        
-        self.gpu_spin = Gtk.SpinButton(adjustment=Gtk.Adjustment(
-            value=10, lower=0, upper=100, step_increment=1
-        ), digits=0, hexpand=True)
-
-        box.append(label)
-        box.append(self.gpu_spin)
-
-        preferences_popover = Gtk.Popover(position=Gtk.PositionType.BOTTOM)
-        preferences_popover.set_child(box)
-        preferences_popover.set_parent(button)
-        preferences_popover.popup()
-        
-    def populate_models(self):
-        try:
-            models = ollama.list()
-        except Exception as e:
-            models = []
-            print ("Failed to fetch models: ", e)
-
-        names = [m.model for m in models["models"]]
-        self.model_store.splice(0, len(list(self.model_store)), names)
-
-    def add_message(self, text="", sender=True):
-        label = Gtk.Label()
-        label.set_selectable(True)
-        label.set_wrap(True)
-        label.set_xalign(1.0 if sender else 0.0)
-        label.set_text(text)
-        set_margin_all(label, 8)
-        
-        row = Gtk.ListBoxRow()
-        row.set_child(label)
-
-        self.listbox.append(row)
-        
-        adj = self.listbox.get_adjustment()
-        adj.set_value(adj.get_upper())
-
-        return label
-
-    def add_response_part(self, part):
-        if self.cur_ai_response == None:
-            return
-
-        self.cur_ai_response.set_label(self.cur_ai_response.get_label() + part)
-        
-    def send_message(self, entry, _=None):
-        buffer = entry.get_buffer()
-        text = buffer.get_text()
-        if text == None or text == "":
-            return
-
-        buffer.delete_text(0, -1) # Delete all text
-        self.add_message(text=text)
-        self.cur_ai_response = self.add_message(sender=False)
-
-        self.listbox.grab_focus()
-        self.entry.set_sensitive(False)
-        
-        app = self.get_application()
-        model = self.model_dropdown.get_selected_item().get_string()
-        num_gpu = self.gpu_spin.get_value_as_int() if self.gpu_spin else 10
-        app.loop.call_soon_threadsafe(lambda: asyncio.create_task(
-            app.send_prompt(model, text, num_gpu)
-        ))
-        
-class PyLlama(Gtk.Application):
-    def __init__(self):
-        super().__init__(application_id="eur.cfpi-fpsi.pyllama")
-        GLib.set_application_name("PyLlama")
-        
-        self.client = ollama.AsyncClient()
-        self.loop = asyncio.new_event_loop()
-        GLib.timeout_add(100, self.run_asyncio_loop)
-        
-    def do_activate(self):
-        self.window = PyLlamaWindow(application=self, title="PyLlama")
-        self.window.present()
-
-    def run_asyncio_loop(self):
-        self.loop.stop()
-        self.loop.run_forever()
-        return True
-    
-    async def send_prompt(self, model, text="", num_gpu=10):
-        messages = [{ "role": "user", "content": text }]
-        GLib.idle_add(self.window.hold_prompt, True)
-        async for part in await self.client.chat(model=model, messages=messages, stream=True, options={ "num_gpu": num_gpu }):
-            GLib.idle_add(self.window.add_response_part, part.message.content)
-        GLib.idle_add(self.window.unhold_prompt, True)
-        
-if __name__ == "__main__":
-    app = PyLlama()
-    app.run()
diff --git a/util.py b/util.py
new file mode 100644
index 0000000..fb6b602
--- /dev/null
+++ b/util.py
@@ -0,0 +1,5 @@
+def set_margin_all(widget, m):
+    widget.set_margin_start(m)
+    widget.set_margin_top(m)
+    widget.set_margin_end(m)
+    widget.set_margin_bottom(m)