i)B-ddlZddlZddlmZmZmZmZmZddlZ ddl m Z eje Z ddhZgdZgdZgdgd gd dZ d&d ed eed eedeefdZ d'dedeedefdZ d(de j.de j.deedfdeeeeefded ed eed eededee j.dedededee j.dee j.d ed!ed"ed#ed$edeeeff*d%Zy))N)DictListOptionalTupleUnion)tqdmeulerheun)?g@@)r .袋????۶m۶m?竪?ى؉??UUUUUU?%I$I??tE]t??皙??333333??qq??)r rrrrrrr)r rrrrrrr)r r rrrrrrshift timesteps infer_stepsreturncd}|t|}|r#|ddk(r|j|r |ddk(rt|dkrtj dnUt|dkDr$tj dt||dd}|Dcgc]}t t |fd }}|}|J|H|dkDrCt|Dcgc] }d ||z z  }}d k7r|Dcgc]}|zd d z |zzz }}|}|;} t tfd  | k7rtj d | t}|Scc}wcc}wcc}w) aCompute the timestep schedule for diffusion sampling. When ``infer_steps`` is provided and ``timesteps`` is None, a continuous linspace schedule is generated (matching the PyTorch base-model behaviour). The legacy lookup-table path (8-step ``SHIFT_TIMESTEPS``) is used only when neither ``timesteps`` nor ``infer_steps`` is supplied. Args: shift: Diffusion timestep shift (applied via ``shift*t / (1+(shift-1)*t)``). timesteps: Optional custom list of timesteps. infer_steps: Number of diffusion steps. When given, overrides the fixed 8-step lookup table. Returns: List of timestep values (descending, without trailing 0). Nrz 20; truncatingct||z SNabs)xts V/mnt/workspace/acestep.cpp/tests/../../ACE-Step-1.5/acestep/models/mlx/dit_generate.pyz'get_timestep_schedule..Ss c!a%jkeyr c t|z Sr)r*)r,r s r.r/z'get_timestep_schedule..^s AIr0z.shift=%.2f rounded to nearest valid shift=%.1f) listpoplenloggerwarningminVALID_TIMESTEPSrange VALID_SHIFTSSHIFT_TIMESTEPS) r r!r"t_schedule_listts_listr-mappediraworiginal_shifts ` r.get_timestep_schedulerD2sh*Oy/'"+* KKM'"+* w? U " NNK^]b c)%0 !\BFsD;E Eguidance_scalemomentum_statenorm_thresholdcddlm}d}||z }|||jddz}||d<|dkDrQ|j||zj |d}|j |j |||dzz } || z}||j||zj |ddzz } || zj |d| z} || z } ||dz | zzS)uAPG (Adaptive Projected Guidance) in pure MLX — mirrors the PyTorch ``apg_forward``. Projection is performed along axis 1 (the time/sequence dimension) to match the PyTorch implementation which calls ``apg_forward(..., dims=[1])``. rNr&runningTaxiskeepdimsg:0yE>)mlx.corecoregetsqrtsumminimum ones_like) pred_cond pred_uncondrErFrGmx proj_axisdiff diff_norm scale_factorv1parallel orthogonals r._mlx_apg_forwardr^fsI { "D!n((A66$(y!GGTD[--9t-LM zz",,y"9>YY]M];^_ l" bggy9499ySW9XY\`` aBr I=BHJ *j8 88r0encoder_hidden_states_npcontext_latents_npsrc_latents_shape.seed infer_methodnull_condition_emb_npcfg_interval_startcfg_interval_endaudio_cover_strength"encoder_hidden_states_non_cover_npcontext_latents_non_cover_np compile_model disable_tqdm sampler_modevelocity_norm_thresholdvelocity_ema_factorc CDEFGHIJddlmGddlm}|tvrt d|dtd|dk(IdkDJdkDHIr0|d k(rt jd nt jd Jrt jd Hrt jd i}tj}Gj|}Gj|}|Gj|nd}|Gj|nd}|dD|d}|d} dkDxr| duEErGj| nd}ErGj||j}Gj||gd}Gj||gd}|1Gj||j} Gj|| gd}|Gj||gd}ErindF|GjjD||f}!nt!|t"rg}"|D]}#|#|#dkr.|"j%Gjjd||f8Gjj't)|#}$|"j%Gjjd||f|$Gj|"d}!nDGjj't)|}$GjjD||f|$}!t+|||}%t-|%}&t)|&| z}'dC|r,fd}( Gj/|(Ct jdIrd}*n C|nd}*|!}+d},CEGIfd}-D E Ffd}.GHJfd}/tj}0d}1t3t5|&d|D]}2|%|2}3|2|'k\r|1sd}1||}|}|*|}*ErGj|+|+gdn|+}4|-|4|3|||*\}5}*Gj7|5|.|5|3}5|/|5|+|,}5|2|&dz k(r0Gj9Dddf|3}6|+|5|6zz }+Gj7|+n*|%|2dz}7Ir|dk(r|3|7z }8Gj9Dddf|8}9|+|5|9zz }:Gj7|:ErGj|:|:gdn|:};|-|;|7|||*\}<}*Gj7|<|.|<|7}<|/|<|:|5}Gjj|+j}?|7|?zd|7z |>zz}+n"|3|7z }8Gj9Dddf|8}9|+|5|9zz }+Gj7|+|5},tj}@tj}A|@|0z |d<|dt;|&dz |d <|A|z |d!<||d"<t=j|+}B|B|d#S#t0$r!})t jd|)Yd})~)d})~)wwxYw)$u:Run the complete MLX diffusion loop with optional CFG guidance. This is the core generation function. It accepts numpy arrays (converted from PyTorch tensors by the handler) and returns numpy arrays that the handler converts back to PyTorch. Args: mlx_decoder: ``MLXDiTDecoder`` instance with loaded weights. encoder_hidden_states_np: [B, enc_L, D] from prepare_condition (numpy). context_latents_np: [B, T, C] from prepare_condition (numpy). src_latents_shape: shape tuple [B, T, 64] for noise generation. seed: random seed (int, list[int], or None). infer_method: "ode" or "sde". shift: timestep shift factor. timesteps: optional custom timestep list. infer_steps: number of diffusion steps. guidance_scale: CFG guidance strength (>1.0 enables CFG). null_condition_emb_np: [1, 1, D] null condition embedding for CFG. cfg_interval_start: timestep ratio below which CFG is disabled. cfg_interval_end: timestep ratio above which CFG is disabled. audio_cover_strength: cover strength (0-1). encoder_hidden_states_non_cover_np: optional [B, enc_L, D] for non-cover. context_latents_non_cover_np: optional [B, T, C] for non-cover. compile_model: If True, compile the decoder step with ``mx.compile``. disable_tqdm: If True, suppress the diffusion progress bar. sampler_mode: Sampler algorithm — ``"euler"`` (first-order, default) or ``"heun"`` (second-order predictor-corrector for cleaner output). velocity_norm_threshold: Clamp velocity prediction L2 norm relative to input norm at each step. 0 disables (default). Values around 1.5–3.0 reduce outlier artefacts. velocity_ema_factor: Blend current velocity prediction with the previous step's prediction via EMA (``vt = (1-f)*vt + f*prev``). 0 disables (default). Values around 0.05–0.2 smooth the trajectory. Returns: Dict with ``"target_latents"`` (numpy) and ``"time_costs"`` dict. rNr&)MLXCrossAttentionCachezUnsupported sampler_mode 'z'. Expected one of .r sdez[MLX-DiT] Heun sampler is not supported with SDE inference method. Falling back to Euler for SDE steps. Use infer_method='ode' for Heun.zF[MLX-DiT] Using Heun (second-order) sampler for higher-quality output.z:[MLX-DiT] Velocity norm clamping enabled (threshold=%.2f).z7[MLX-DiT] Velocity EMA smoothing enabled (factor=%.3f).r )rKr1)r"c ,|||||dd\}}|S)NF hidden_statestimestep timestep_rencoder_hidden_statescontext_latentscache use_cache)xtr-trencctxvt_ mlx_decoders r. _raw_stepz)mlx_generate_diffusion.._raw_steps) 1&)3eEB Ir0z4[MLX-DiT] Diffusion step compiled with mx.compile().z:[MLX-DiT] mx.compile() failed (%s); using uncompiled path.c  j|jdf|}||||||fS |||||| xr \}}||fS)zSingle model evaluation helper.rru)fullshape) x_inputt_valrctx_in step_cachet_arrvt_out_compiled_stepdo_cfgrrVuse_heuns r. _model_evalz+mlx_generate_diffusion.._model_eval(swq)+U3  %!'5%fEzQ Q(!"%"!z2(l  z!!r0c^s|S|d}|d}|cxkrkrn|St|| S|S)zApply CFG guidance if enabled.N)r^) vt_raw current_t_valrTrUbszrfrerrErFs r. _apply_cfgz*mlx_generate_diffusion.._apply_cfg8sNM4CL STl  B2B B$I{NN[ [r0c(r|j||zjdd}j||zjdddz}jj| |z|dzz }||z}r|d z |z |zz}|S)z/Apply optional norm clamping and EMA smoothing.)r&rsTrJg|=r )rPrQrRrS) vt_guided xt_current prev_velocityvt_normxt_normscalerVuse_emause_norm_clamprnrms r._apply_stabilisationz4mlx_generate_diffusion.._apply_stabilisationBs ggy9499vPT9UVGggzJ6;;RV;WX[``GJJ W%(72wGE"E)I }022i?BUXeBeeIr0FzMLX DiT diffusion)descdisableToderdiffusion_time_costdiffusion_per_step_time_costtotal_time_costrl)target_latents time_costs)rMrN dit_modelrpVALID_SAMPLER_MODES ValueErrorr7r8infotimearray broadcast_tor concatenaterandomnormal isinstancer4appendr2intrDr6compile Exceptionrr;evalrmaxnp)Krr_r`rarbrcr r!r"rErdrerfrgrhrirjrkrlrmrnrpr total_startenc_hsr enc_hs_ncctx_ncTC null_cond null_expandednull_expanded_ncnoisepartssr2r> num_steps cover_stepsrexcr{r~prev_vtrrr diff_start_switched_to_non_coverstep_idx current_tx_inrt_unsqnext_tdtdt_arr xt_predictedx_in2vt2vt_avg pred_clean new_noisediff_end total_end result_nprrrrFrVrrrsK` ` `` `` @@@@@@@@r.mlx_generate_diffusionrsx1..5l^CVWjVkklmnnv%H,q0N!A%G 5 NNX  KK` a PRij MObcJ))+K XX. /F ((% &C@b@n;rs:( 55   8 $'  < K (  $!%1 1~1#1 %[ 1p&* 99TN 9  9L-1 $!%26 #!"%?C9=%(!$+\ jj\ \S#X \ 5d3i( ) \  \ \~\#\\$BJJ/\\\ \)1(<\ #+2::"6!\"#\$%\&'\(#)\*+\, #v+-\r0