ni{%SrSSKrSSKrSSKrSSKJr SSKJrJrJ r J r SSK J r Sr SrSrS rS rS \- rS rS rSr1SkrS\4SjrS\4SjrS\4SjrS\4SjrS\4SjrSS\S-S\4SjjrSS\S\\ \\44SjjrSS\S\4Sjjr SSSSSSSS S.r!SS S!S".S#S S$S".S%S&S'S".S(.r"S S!S S$S).r#S r$S*\%S+\%S\4S,jr&S-\%S\%4S.jr'\"S/S055r(S1\(S\(4S2jr)SS3\\%S4\S0S\%4S5jjr*S6S7S8S8S9/S:S;S.S?S@S8S8S9/S:S;S.S?S@SASAS=SB/SBS;S.S?S@SASES=SB/SBS;S.S?S@SESES=SBSF/SFS;S.S?S@SESGS=SBSF/SFS;S.S?S?SGSGS=/SIQSFS;S.S@S@SGSGS=/SIQSKS;S.SL.r+\+SM\+SN'S\,4SOjr-SPr.SQ\,S\%4SRjr/SSS\\,S\4STjjr0SSQ\\,S\(4SUjjr1SV\%S\%4SWjr2SX\%SY\\%S\4SZjr3S[\%S\\\%S\\%4S]jr4SV\%S^\,S\ \,\,44S_jr5SS`\,Sa\%S\(4Sbjjr6SS\S\,4Scjjr7S\,4Sdjr8SSe\Sf\,Sa\%Sg\Sh\%S\,4 Sijjr9Sj\,S4\(Sk\S\ \\%44Sljr:Se\S4\(Sk\S\ \\%44Smjr;SV\%S4\(S\ \\%44SnjrSqSrSsStSuSvSwSxSL.r?\@"\?R55rBS\%4SyjrCSz\%S\(4S{jrDSqE\\(\FS|'S\(4S}jrGS1\(4S~jrHg)a GPU Configuration Module Centralized GPU memory detection and adaptive configuration management Debug Mode: Set environment variable MAX_CUDA_VRAM to simulate different GPU memory sizes. Example: MAX_CUDA_VRAM=8 python acestep # Simulates 8GB GPU For MPS testing, use MAX_MPS_VRAM to simulate MPS memory. Example: MAX_MPS_VRAM=16 python acestep # Simulates 16GB MPS This is useful for testing GPU tier configurations on high-end hardware. N) dataclass)OptionalListDictTuple)logger MAX_CUDA_VRAM MAX_MPS_VRAM MAX_XPU_VRAMACESTEP_SAVE_MEMORY?0@g4@z&https://download.pytorch.org/whl/cu121z(https://download.pytorch.org/whl/rocm6.0>ptmlxvllmreturnc[RS:wagSSKn[URS5=(a$ URR R 5$![a gf=f)zCheck if running on macOS with MPS (Apple Silicon) available. This is the canonical check used across the codebase to apply Mac-specific configuration overrides (no compile, no quantization, mlx backend, no offload, etc.). darwinFrNmps)sysplatformtorchhasattrbackendsr is_available Exceptionrs 1/mnt/workspace/ACE-Step-1.5/acestep/gpu_config.pyis_mps_platformr,sV ||xu~~u-S%..2D2D2Q2Q2SS sAA A)(A)c`SSKnURR5$![a gf=f)z)Return whether CUDA runtime is available.rNF)rcudarrrs ris_cuda_availabler"=s/zz&&(( s  --cSSKn[URS5=(a$ URRR 5$![ a gf=f)z(Return whether MPS runtime is available.rNrF)rrrrrrrs ris_mps_availabler$GsEu~~u-S%..2D2D2Q2Q2SS sAA AAcSSKn[US5=(a URR5$![a gf=f)z(Return whether XPU runtime is available.rNxpuF)rrr&rrrs ris_xpu_availabler'Qs;ue$A)?)?)AA s03 AAcSSKnURR5=(a5 [URS5=(a URR SL$![ a gf=f)aReturn whether the active CUDA device is an AMD ROCm/HIP device. On ROCm, PyTorch exposes the GPU as a CUDA device but also sets ``torch.version.hip``. This function returns ``True`` only when *both* conditions hold: a CUDA device is present **and** the build is a ROCm/HIP build. rNhipF)rr!rrversionr)rrs ris_rocm_availabler+[s_  JJ # # % . u- . !!- sAA A)(A) device_indexcSSKnURR5(dgURRU5up#US:$![a gf=f)z>Return whether a CUDA device supports native bfloat16 kernels.rNFrr!rget_device_capabilityr)r,rmajor_s rcuda_supports_bfloat16r3osRzz&&((::33LAz s#A!A AAcSSKnURR5(aURRU5$g![a gf=f)z>Return the active CUDA device capability tuple when available.rNr/)r,rs rget_cuda_device_capabilityr5|sP :: " " $ $::33LA A %  s=A AAcj[U5nUSL=(a USS:=(a [5(+$)zDReturn True for pre-Volta CUDA GPUs that should avoid vLLM defaults.Nr)r5r+)r, capabilitys ris_legacy_cuda_gpur9s1+L9J T ! Sjma&7 S@Q@S "xl_turbo" "acestep-v15-xl-base" -> "xl_base" "acestep-v15-xl-sft" -> "xl_base" (sft shares base VRAM profile) "acestep-v15-turbo" -> "turbo" "acestep-v15-base" -> "base" "acestep-v15-sft" -> "base" (sft shares base VRAM profile) xlrOrPxl_)lowerrX)rYrTis_xlvariants rget_dit_type_from_pathrasO  2 $ $ &D D$ 'Ew%%#S ?00r:c\rSrSr%Sr\\S'\\S'\\S'\\S'\\S'\\S'\ \S '\ \\S '\\S '\\S '\\S '\ \S'\ \S'\ \S'\ \S'\ \\4\S'Sr \ \S'Sr \\S'Srg) GPUConfigz+GPU configuration based on available memorytier gpu_memory_gbmax_duration_with_lmmax_duration_without_lmmax_batch_size_with_lmmax_batch_size_without_lminit_lm_defaultavailable_lm_modelsrecommended_lm_modellm_backend_restrictionrecommended_backendoffload_to_cpu_defaultoffload_dit_to_cpu_defaultquantization_defaultcompile_model_default lm_memory_gbFsave_memory_modemlx_vae_chunk_sizeN)__name__ __module__ __qualname____firstlineno____doc__str__annotations__floatintboolrrrurw__static_attributes__rxr:rrcrcs5 I   ""c"   !   sEz"" #d""!r:rcconfigcl[5(a$[R"S5 SUlSUlU$)z:Apply runtime hardware overrides for LM backend selection.zjLegacy CUDA GPU detected (pre-Volta compute capability): forcing 5Hz LM backend recommendation to PyTorch.pt_onlyr)r9rinfornrors r)_apply_lm_backend_compatibility_overridesrs4  @ )2%%)" Mr:requested_backend gpu_configcU=(d [5n[USS5n[USS5nU=(d SR5R5nU[;aUnU[;aSnUS:XagUS:XaUS:Xa UnUS ;aSnU$U$) zBResolve the LM backend against runtime compatibility restrictions.rorrnallr[rr pt_mlx_only>rr)get_global_gpu_configgetattrstripr^VALID_LM_BACKENDS)rrrrornbackendfallbacks rresolve_lm_backendrs  202F!&*?H$V-EuM &B--/557G''% + +G*.7f3D& = (H Nr:hFr[rrT)rgrhrirjrkrlrmrnrorprqrrrsrtXacestep-5Hz-lm-0.6BrLacestep-5Hz-lm-1.7Br.)rLrM)rracestep-5Hz-lm-4B r)tier1tier2tier3tier4tier5tier6atier6b unlimitedrtier6c  [RR[5nUGb[ U5n[ R "SUSS[S35 SSKnURR5(aURRS5RnUS- nX:a[RSS 5n[S X- 5nXd- n[S [S U55nURRU5 [ R "S US SUSSUSSUSSUSS3 5 U$[RR[$5n U b0[ U 5n[ R "SUSS[$S35 U$[RR[&5n U b0[ U 5n[ R "SUSS[&S35 U$SSKnURR5(aURRS5Rn U S- n URR)S5n [+UR,S5=(a UR,R.SLnU(a7[ R0"SU SU SSUR,R.S35 U $[ R0"SU SU SS 35 U $[+US!5(aUR2R5(apUR2RS5Rn U S- n [5UR2RS5S"S#5n [ R0"S$U SU SS 35 U $[+UR6S%5(aUR6R8R5(a[5US%S5nUb([+US&5(aUR;5n U S- n U $Ub=[+US'5(a,URS5n[5US(S5n U (aU S- n U $SSKnUR?/S*QS+S+S,S-9n[AURBRE55nUS- S.-n U $[GU5 g![ a$n[ R "SU35 SnAU$SnAff=f!["a% [ R "S[SUS35 GNf=f!["a% [ R "S[$SU S35 GNf=f!["a% [ R "S[&SU S35 GNnf=f![ a$n[ R "S)U35 SnAGNBSnAff=f![ a! [ R "S/[$S035 g1f=f![ a#n[ R "S2U35 SnAgSnAff=f)3a Get GPU memory in GB. Returns 0 if no GPU is available. Debug Mode: Set environment variable MAX_CUDA_VRAM to override the detected GPU memory. Example: MAX_CUDA_VRAM=8 python acestep # Simulates 8GB GPU For MPS testing, set MAX_MPS_VRAM to override MPS memory detection. Example: MAX_MPS_VRAM=16 python acestep # Simulates 16GB MPS This allows testing different GPU tier configurations on high-end hardware. Nu,⚠️ DEBUG MODE: Simulating GPU memory as .1fz GB (set via z environment variable)r@rCr r;rHu/⚠️ DEBUG MODE: Set CUDA memory fraction to z.4fz (allocator_budget=.2fzGB, ref_context=z GB, target=z GB, total=zGB) to enforce hard VRAM capu6⚠️ DEBUG MODE: Could not enforce CUDA memory cap: zInvalid z value: z , ignoringu,⚠️ DEBUG MODE: Simulating MPS memory as u,⚠️ DEBUG MODE: Simulating XPU memory as r)zROCm GPU detected:  (z GB, HIP )zCUDA GPU detected: z GB)r&name Intel XPUzIntel XPU detected: rrecommended_max_memoryget_device_properties total_memoryzFailed to detect MPS memory: )sysctlz-nz hw.memsizeT)capture_outputtexttimeoutg?z0MPS available but total memory not exposed. Set z to enable tiering.rIzFailed to detect GPU memory: )$osenvirongetDEBUG_MAX_CUDA_VRAM_ENVrrwarningrr!rrr MODEL_VRAMmaxminset_per_process_memory_fractionr ValueErrorDEBUG_MAX_MPS_VRAM_ENVDEBUG_MAX_XPU_VRAM_ENVget_device_namerr*r)rr&rrrr subprocessrunrstdoutr_log_gpu_diagnostic_info) debug_vram simulated_gbr total_bytestotal_gbREFERENCE_CONTEXT_GBallocator_budget_gbfractionedebug_mps_vramdebug_xpu_vramr memory_gb device_nameis_rocm mps_modulepropsrresulttotal_system_bytess rget_gpu_memory_gbrs 78J,  ,L NN>|C>PP\]t\uvLM  ]::**,,"'**"B"B1"E"R"RK*g6H#.0:~~nc/R,.1!D/+$7#A#&tSh-?#@ BB8LMhWZ^\11DS0IJ++?*DKP\]`Oab%%-cN2NP  ZZ^^$:;N!  0L NN>|C>PP\]s\tuKL  ZZ^^$:;N!  0L NN>|C>PP\]s\tuKL   G :: " " $ $ ::;;A>KKL$0I**44Q7KemmU3U 8I8IQU8UG )+b3yQVQ^Q^QbQbPccde   1+b3tTU  UE " "uyy'='='?'? 99::1=JJL$0I! //2FKK KK.{m2i_DQ R  U^^U + +0B0B0O0O0Q0Q t4J D)g 8//$.#D#D#FL , 8I$$)g 7//'<%?"/7;tC    %U +M ]!WXYWZ[\\  ]  NN238J:J*U    NN12(>:J*U  ^ D!>qcBCC D" FG]F^^qr   6qc:;s-R7CR-R7.S),.TCV,7V,BV,7AV, *U 4?U 5AU>: V, R4R/)R7/R44R77+S&%S&)+TT+U  U  U;U60V,6U;;V,>(V)&V,(V))V,, W6WWc [R"S5 [R"S5 [R"S5 [URS5=(a URRSLn[URS5=(a URR SLnU(Ga[R"S5 [R"SURR35 [R"S5 [R"S 5 [R"S5 [R"S 5 [R"S 5 [R"S 5 [R"S 5 [R"S5 [R"S5 [ RRS5nU(a[R"SU35 On[R"S5 [R"S5 [R"S5 [R"S5 [R"S5 [R"S5 [R"S5 [R"S5 [R"S5 [R"S5 [R"S5 [R"S5 [R"S5 [R"S5 [R"S5 GOU(Ga[R"S5 [R"S URR 35 [R"S5 [R"S 5 [R"S5 [R"S!5 [R"S"5 [R"S#5 [R"S$5 [R"S5 [R"S5 [R"S%5 [R"S&5 [R"S'5 [R"S(5 [R"S)[35 GO[R"S*5 [R"S5 [R"S+5 [R"S5 [R"S,5 [R"S-[35 [R"S5 [R"S.5 [R"S/5 [R"S0[35 [R"S5 [R"S15 [R"S5 g)2z Log diagnostic information when GPU is not detected to help users troubleshoot. Args: torch_module: The torch module to inspect for build information zP================================================================================u0⚠️ GPU NOT DETECTED - DIAGNOSTIC INFORMATIONr)Nr!u✓ PyTorch ROCm build detectedz HIP version: r[u,❌ torch.cuda.is_available() returned Falsez Common causes for AMD/ROCm GPUs:z: 1. ROCm drivers not installed or not properly configuredz0 2. GPU not supported by installed ROCm versionzG 3. Missing or incorrect HSA_OVERRIDE_GFX_VERSION environment variablez. 4. ROCm runtime libraries not in system pathHSA_OVERRIDE_GFX_VERSIONz& HSA_OVERRIDE_GFX_VERSION is set to: u, ⚠️ HSA_OVERRIDE_GFX_VERSION is not setz5 For RDNA3 GPUs (RX 7000 series, RX 9000 series):zH - RX 7900 XT/XTX, RX 9070 XT: set HSA_OVERRIDE_GFX_VERSION=11.0.0zD - RX 7800 XT, RX 7700 XT: set HSA_OVERRIDE_GFX_VERSION=11.0.1z5 - RX 7600: set HSA_OVERRIDE_GFX_VERSION=11.0.2zTroubleshooting steps:z 1. Verify ROCm installation:z% rocm-smi # Should list your GPUz 2. Check PyTorch ROCm build:zB python -c "import torch; print(f'ROCm: {torch.version.hip}')"z: 3. Set HSA_OVERRIDE_GFX_VERSION for your GPU (see above)zJ 4. On Windows: Use start_gradio_ui_rocm.bat which sets required env varszA 5. See docs/en/ACE-Step1.5-Rocm-Manual-Linux.md for Linux setupzB 6. See requirements-rocm.txt for Windows ROCm setup instructionsu✓ PyTorch CUDA build detectedz CUDA version: zCommon causes for NVIDIA GPUs:z! 1. NVIDIA drivers not installedz3 2. CUDA runtime not installed or version mismatchz0 3. GPU not supported by installed CUDA versionz' 1. Verify NVIDIA driver installation:z' nvidia-smi # Should list your GPUz% 2. Check CUDA version compatibilityz) 3. Reinstall PyTorch with CUDA support:z# pip install torch --index-url u#⚠️ PyTorch build type: CPU-onlyz1You have installed a CPU-only version of PyTorch!zFor NVIDIA GPUs:z pip install torch --index-url zFor AMD GPUs with ROCm:z> Windows: See requirements-rocm.txt for detailed instructionsz' Linux: pip install torch --index-url z=For more information, see README.md section 'AMD / ROCm GPUs') rrrr*r)r!rrrPYTORCH_CUDA_INSTALL_URLPYTORCH_ROCM_INSTALL_URL) torch_module is_rocm_build is_cuda_build hsa_overrides rrr~s NN8 NNEF NN8  $$e,U1E1E1I1IQU1U  $$f-W,2F2F2K2KSW2W89)=)=)A)A(BCDrEFr9:STIJ U  GHrzz~~&@A  NNCL>R S NNI J NNR S NNZ  NNV  NNR Sr/078>?78 R  ST X   O   P  89),*>*>*C*C)DEFrEFr78:;LMIJr/0@A@A>?BC<=U   sA// A<;A<c Uc [5n[U5n[Un[5nU(a[R "SUSSUS35 [ S0SU_SU_SUS_SUS_S US _S US _S US _S US _S URS S5_SU(aSOURSS5_SU(aSOURSS5_SU(aSOURSS5_SU(aSOURSS5_SU(aSOURSS5_SU(aSOURSS5_SUS_SU(a [U5OS_6n[U5$)u Get GPU configuration based on detected or provided GPU memory. On macOS with MPS (Apple Silicon), several overrides are applied automatically regardless of the tier selected by memory size: - ``compile_model_default = False`` — ``torch.compile`` is not supported on MPS and would error or silently fall back to eager mode. - ``quantization_default = False`` — torchao INT8 quantization is incompatible with MPS / macOS. - ``recommended_backend = "mlx"`` — MLX provides native Apple Silicon acceleration for the 5Hz LM; vllm requires CUDA. - ``lm_backend_restriction = "pt_mlx_only"`` — vllm cannot run on MPS. - ``offload_to_cpu_default = False`` — Apple Silicon uses unified memory; offloading to CPU provides no benefit and adds overhead. - ``offload_dit_to_cpu_default = False`` — same reason. Args: gpu_memory_gb: GPU memory in GB. If None, will be auto-detected. Returns: GPUConfig object with all configuration parameters zmacOS MPS detected (rz GB unified memory, tier=zb). Applying Apple Silicon optimizations: no compile, no quantization, mlx backend, no CPU offload.rerfrgrhrirjrkrlrmr[rnrrrorrrpFTrqrrrsrtrwrvrx) rrGPU_TIER_CONFIGSrrrrcrrr)rfrer_mpss rget_gpu_configr-s0)+  &D d #F  D  "="55NtfU+ + & &#&$$:;&!''@ A & &&>? & #))D"E &01&##89&$ZZ(>C&  - ZZ0% 8& " ZZ-v 6!&&  % ZZ0$ 7)&, $) ZZ4d ;/&4 # ZZ. 57&> $ ZZ/ 6A&BN+C&H 4MB K&FN 5V <2d2Qs A5 A5total_gpu_memory_gbc"[U5n[RU[S5nUSnUSnUnXE-S-nSnSSKn U RR 5(Ga(U RR 5upU S- nU S- n [RR[5n U bc[U 5nX:aR[RSS 5n[S X- 5nU RR5S- n[SUU- 5nUn S n[SUU- [ - 5n[#UU5nX- nUU-nUU - n[#S [S U55n[$R&"S USUSSUSSUSSUSSUS3 5 UU4$US:a[#S [SXq- 55nUU4$[#S [S Xq- 55nUU4$![[4a Nf=f![(a$n[$R*"SUS35 SnAN}SnAff=f)a) Calculate GPU memory utilization ratio for LM model. This function now uses *actually free* VRAM (via torch.cuda.mem_get_info) when available, instead of computing the ratio purely from total VRAM. This is critical because DiT, VAE, and text encoder are already loaded when the LM initializes, so the "available" memory is much less than total. Args: model_path: LM model path (e.g., "acestep-5Hz-lm-0.6B") total_gpu_memory_gb: Total GPU memory in GB (used as fallback) Returns: Tuple of (gpu_memory_utilization_ratio, target_memory_gb) rLrErGrDNrrrCr g?g?g?z [get_lm_gpu_memory_ratio] model=z, free=rzGB, current_usage=zGB, lm_target=zGB, usable_for_lm=z GB, ratio=z.3fz5[get_lm_gpu_memory_ratio] Failed to query free VRAM: z, using fallbackrg?)rLM_VRAMrrr!r mem_get_inforrrrrrmemory_reservedr TypeErrorVRAM_SAFETY_MARGIN_GBrrrrr)rr rlm_info lm_weights_gblm_kv_cache_gb target_gbtotal_target_gbfree_gbr free_bytesractual_total_gbrrref_context_gbr reserved_gbdit_reserve_gb usable_for_lmcurrent_usage_gbdesired_total_usageratiors rget_lm_gpu_memory_ratior#so$#:.Jkk*gfo6GI&M]+NI#4s:OG4  :: " " $ $&+jj&=&=&? #J G,G)W5O(?@J% #(#4L#5)3)L.1#|7T.U+&+jj&@&@&Bg&N "%a)<{)J"K*6!N7^#;>S#STM ?M /8 "2]"B '/9ESe_-E KK2:,ggc]S!!1# 6n_UXDYZ!!.s 3:eC[J  )# #[ %hb CS/"GHI ) CS/"GHI ) O#I.8  CA3FV W   s>A0G 0A"G A8G GG GG H*H  H total_vram_gbdit_typec SU3[;aSU3OSn[U[S-[S-[S-[S-[-nX- nUS::a [U5$/n0nSHZn[UnUS US -n [R US 5n X-U::dM7S U3n UR U 5 US US -Xg'M\ [R US 5n [S [XJ- 55n [U S5n U(aH[UR55Sn XmnXN- n[S [X- 55n[US5nOU nUS:aSnSnO0US:aSnSnO%US:aSnSnOUS:aSnSnOUS:aSnSnOSnSn[U5n[R U05n[S-0SU_SU_SU_SU_SU_SU _S[U5_S U_S!UR S!U(aUSOS"5_S#UR S#S$5_S%UR S%S&5_S'UR S'S(5_S)UR S)S(5_S*UR S*S(5_S+UR S+S(5_S,U_6n[!U5$).a Compute GPU configuration based on what actually fits in VRAM. This is a VRAM-budget-based approach: instead of hard-coded tier boundaries, we calculate how much memory each component needs and determine what fits. Args: total_vram_gb: Total GPU VRAM in GB dit_type: DiT type key -- "turbo", "base", "xl_turbo", "xl_base", etc. (affects model weight size and inference VRAM due to CFG) Returns: GPUConfig with parameters that fit within the VRAM budget dit_r<r@rArCrBrrKrErGrJzacestep-5Hz-lm-rr.rrrrrrrrerfrgrhrirjrkrlrmr[rnrrorrpTrqrrrsrtrx)rrrrDIT_INFERENCE_VRAM_PER_BATCHrappendrrrlistkeysrrrcrr)r$r%dit_key base_usage availablerlrtsize_keyrlm_totalinference_per_batch model_namemax_batch_no_lmlargest_lm_sizelm_usageremaining_for_inferencemax_batch_with_lm max_dur_lm max_dur_no_lmre tier_configrs rcompute_adaptive_configr>.s< &*(#4 #BXJ G7 U   ^ $ % ^ $ % % &  '  *IA~m,,L*(#9% (>>:>>xM  )Y 6*8*5J  & &z 2%,Y%7'-:P%PL "+7::8SI!S!@ABO/1-O|0023B70"+"63'>'T#UV 115+  "   "   "   !      &D"&&tR0K  #(!.  1  #2 010)__ "&9  "r  +/GO(OO,A6J +/GN $/??3OQU#V!")__-CTJ#$*oo.EtL%&"'F* 5V <R(i89++7. <  k!! r:c U(a URO URnX:a*SUSURSSUSU(aSOSSUS 3 nS U4$g ) a Check if requested batch size is within limits for current GPU configuration. Args: batch_size: Requested batch size gpu_config: Current GPU configuration lm_initialized: Whether LM is initialized Returns: Tuple of (is_valid, warning_message) u⚠️ Requested batch size (z") exceeds the limit for your GPU (rrYrr[r\z$ LM). Batch size will be clamped to .Fr])rirjrf)rLrrWmax_batch_sizer_s rcheck_batch_size_limitrdEs  ))  1 1 "+J<8((--CNCST(i89--;,E F  #:.J&99  (:  :,b 5  $ $S ))@:KiKiAj@k m r:cDUR(dgURS$)z Get recommended LM model for current GPU configuration. Args: gpu_config: Current GPU configuration Returns: Recommended LM model path, or None if LM is not supported Nr()rlrs rget_recommended_lm_modelrks#  ) )  ) )" --r:c[R"S5 [R"SURSS35 [R"SUR35 [R"SURSURS-S 35 [R"S UR SUR S-S 35 [R"S UR 35 [R"S UR35 [R"S UR35 [R"SUR=(d S35 g)z2Print GPU configuration information for debugging.zGPU Configuration:z - GPU Memory: rz GBz - Tier: z - Max Duration (with LM): rZ<z min)z - Max Duration (without LM): z - Max Batch Size (with LM): z! - Max Batch Size (without LM): z - Init LM by Default: z - Available LM Models: NoneN) rrrfrergrhrirjrkrlrjs rprint_gpu_config_inforos) KK$& KK":#;#;C"@DE KK*Z__-./ KK &z'F'F&Gs:KjKjnpKpJqqvw KK )**L*L)MSQ[QsQswyQyPzz A KK01R1R0STU KK +J,P,P+QR KK*:+E+E*FGH KK+J,J,J,Tf+UVWr:utier1 (≤4GB)z tier2 (4-6GB)z tier3 (6-8GB)ztier4 (8-12GB)ztier5 (12-16GB)ztier6a (16-20GB)ztier6b (20-24GB)uunlimited (≥24GB)c@SSKnURR5(aURRS5$[ US5(aGUR R5(a(UR R S5n[USS5$[ URS5(aLURRR5(a#SSK nUR5=(d SnUS3$g ![a g f=f![a g f=f) z Get the GPU device name string. Returns: Human-readable GPU name, e.g. "NVIDIA GeForce RTX 4060 Ti", "Apple M2 Pro (MPS)", "CPU only", etc. rNr&rrrz Apple Siliconz (MPS)zApple Silicon (MPS)zCPU onlyzUnknown (PyTorch not available))rr!rrrr&rrrrr processorr ImportError)rrrchips rget_gpu_device_namerts1 :: " " $ $::--a0 0 UE " "uyy'='='?'?II33A6E5&+6 6 U^^U + +0B0B0O0O0Q0Q -))+>v& -, - 101s==DADAD!D D  D D  D DDrec U[;a$[R"SUS35 [5$[ 5n[Un[ 5nU(a[R "SUS35 [S0SU_SU_SUS_SUS_S US _S US _S US _S US _S URS S5_SU(aSOURSS5_SU(aSOURSS5_SU(aSOURSS5_SU(aSOURSS5_SU(aSOURSS5_SU(aSOURSS5_SUS_SU(a [U5OS_6n[U5$)a Create a GPUConfig for a specific tier, applying platform overrides. This is used when the user manually selects a different tier in the UI. The actual gpu_memory_gb is preserved from the real hardware detection, but all tier-based settings come from the selected tier's config. Args: tier: Tier key, e.g. "tier3", "tier6a", "unlimited" Returns: GPUConfig with the selected tier's settings zUnknown tier 'z'', falling back to auto-detected configzManual tier override to u2 on macOS MPS — applying Apple Silicon overridesrerfrgrhrirjrkrlrmr[rnrrrorrrpFTrqrrrsrtrwrvrx) rrrrrrrrcrrr)rereal_gpu_memoryrrs rget_gpu_config_for_tierrws ##v-TUV()O d #F  D  &tf,^ _    % $$:; !''@ A  &&>?  #))D"E 01 ##89 $ZZ(>C   - ZZ0% 8  " ZZ-v 6 "  % ZZ0$ 7% ( $) ZZ4d ;+ . # ZZ. 51 4 $ ZZ/ 67 8N+9 < 4OD ? FB 5V <rIrKrUr`rdrhrkroGPU_TIER_LABELSr-itemsGPU_TIER_CHOICESrtrwrxrrrrxr:rrsZ  !..*'''00"&DE)"4$$4( t t  S %S/9R TSTT   C DC D#c B    L3LcLdL1116 1"1" 1"h i I )-}% F!$#&"#%& ! ""'%"&&* $!%!,!$#&"#%& ! ""'%"&&* $!%#.!$#&"#%& 56 5"'%"&&* $!% #,!$#&"#%& 56 5"'%"&&* $!% !*!$#&"#%& 57LM 5"'%"&&+ $!%!"A.!,!$#&"#%& 57LM 5"'%"&&+ $!%!"A.#,!$#&"#%& !6"'%"'&+ %!%!"AR8).!$#&"#%& !4"'%"'&+ %!%!"AR8%MZz-X6_5_DfR!!#!HXe_DN=(5/N=YN=b##*15c 2!3!3)-c!3 c]!3H[[*/[ 5%<[|l=5l=Cl=il=^NSNNb(u( )=)=)=)= )=  )=  )=X!*<@ 49@!*<@ 49@cyU4QT9EU<..x}."XiX*      & --/01S1>==#==)==B+/HY'. y   ) r: