perf(#1624): optimizer ladder + COW clone (fixed) + streaming param setter + lazy foundation-scale construct

Directory.Packages.props

@@ -154,16 +154,44 @@
          gated on: Dense -> LayerNorm -> GELU -> Dense now routes its between-matmul ops through the
          FP16-native path instead of the eager FP32 fallback. 0.96.0 supersedes master's 0.95.2;
          Native packages coreleased in lockstep at 0.96.0. -->
-    <!-- 0.96.3: carried the FP16 non-Adam fused-optimizer fix —
-         FusedOptimizerIntegrationTests.Fp16Activations_NonAdamFusedOptimizer_DescendsLoss
-         (FP16 RMSprop) was a training no-op (bit-identical first/last loss) on 0.96.1; fixed there. -->
-    <!-- 0.97.2: published (from master); supersedes 0.96.3.
-         Native packages coreleased in lockstep at 0.97.2. -->
-    <!-- 0.102.3: latest published; carries #658 — forward-GEMM core saturation
-         default-on (s_forwardPackBothBlocking + s_singleRegion, opt-out
-         AIDOTNET_GEMM_FORWARD_PACKBOTH=0 / AIDOTNET_GEMM_SINGLE_REGION=0; ~1.3-2x on
-         transformer/training GEMM shapes) + net471 PackAOnly non-4x4 tail bugfix.
-         Native packages coreleased in lockstep at 0.102.3. -->
+    <!-- 0.101.2: latest published; supersedes 0.98.0. Pulls in the #1624-class OOM/perf work
+         (COW clone, byte-budgeted arena, fused-path per-step arena, streaming clean/dirty eviction,
+         activation-aware autotuner) that this PR's training-scale fixes build on.
+
+         Measured 0.98.0 -> 0.101.2 on the canonical #1624 leak shape (SimCSE<float> dim=384,
+         10 layers, fused-optimizer path under one never-Reset() outer TensorArena, CPU, 31 steps),
+         same local src, version the only variable:
+           * Peak RSS      1404 MiB -> 949 MiB   (-32%, the number that OOMs the 16 GiB runner)
+           * Total alloc   2422 MiB -> 1173 MiB  (-52%; alloc/step 78 -> 38 MiB)
+           * Per-step heap growth flat in BOTH (-0.001 MiB/step: the per-step leak fix is consumer
+             src, constant across the runs), and loss converges identically (1.27 -> 0.26) — the
+             reduction is not from silently breaking training.
+         0.101.2 -> 0.101.5: activation-checkpointing correctness, two fixes.
+         (1) #643 (0.101.4): GradientCheckpointing.Checkpoint's recompute differentiated only w.r.t.
+             the segment input, silently dropping every checkpointed layer's WEIGHT gradients.
+         (2) #645 (0.101.5): multi-segment recompute double-counted earlier segments — a later
+             segment's recompute followed its (non-detached) input back into the earlier segment's
+             checkpoint node, so earlier-segment gradients came out 2x. 0.101.5 detaches each
+             segment's input (matches torch.utils.checkpoint), making sqrt(N) multi-segment exact.
+         Both are required for G4 (#1624) in the diffusion NoisePredictors (sqrt(N) checkpointing) and
+         for NeuralNetworkBase checkpointing.
+
+         0.101.5 -> 0.101.7: the merged MobileNetV3 batch=1 timeout work and supporting fixes —
+         #639 batch=1 BatchNormAffine fusion (#641, required by BatchNormalizationLayer's batch=1
+         training fallback, which calls Engine.BatchNormAffine), Conv2D core-saturation (#647), the
+         nested-arena escaped-buffer #1221 fix (#648), the GPU resident-weight version-gate (#649),
+         and GPU conv-kernel arg-binding fixes (#644).
+         Native packages (OneDNN/OpenBLAS/CLBlast) coreleased in lockstep.
+         Bumped 0.101.7 -> 0.102.2: brings the #632 compiled-inference MemoryPlanning aliasing fix
+         (BlasBatch hoist vs memory planning in attention inference — the root cause of the diffusion
+         Predict_ShouldBeDeterministic failures, which reproduced on 0.101.7 and is gone on 0.102.x),
+         plus the Phase C fp16-weight / fp32-master mixed-precision storage (#650).
+
+         Bumped 0.102.2 -> 0.102.3 (from master, #658): forward-GEMM core saturation default-on
+         (s_forwardPackBothBlocking + s_singleRegion, opt-out AIDOTNET_GEMM_FORWARD_PACKBOTH=0 /
+         AIDOTNET_GEMM_SINGLE_REGION=0; ~1.3-2x on transformer/training GEMM shapes) + the net471
+         PackAOnly non-4x4 scalar-tail bugfix. 0.102.3 is a linear superset of 0.102.2, so the #632 /
+         Phase C deps above are retained. -->
     <PackageVersion Include="AiDotNet.Tensors" Version="0.102.3" />
     <PackageVersion Include="AiDotNet.Native.OneDNN" Version="0.102.3" />
     <PackageVersion Include="AiDotNet.Native.OpenBLAS" Version="0.102.3" />

src/Diffusion/Audio/AudioLDM2Model.cs

@@ -536,7 +536,7 @@ private Tensor<T> GenerateWithDualEncoders(
         }
 
         // Initialize random latent
-        var rng = seed.HasValue ? RandomHelper.CreateSeededRandom(seed.Value) : RandomGenerator;
+        var rng = CreateInferenceRng(seed);
         var latent = SampleNoiseTensor(latentShape, rng);
 
         // Set up scheduler
@@ -689,7 +689,7 @@ public virtual Tensor<T> TransformAudio(
         var startTimestep = Scheduler.Timesteps.Skip(startStep).FirstOrDefault();
 
         // Add noise at starting timestep
-        var rng = seed.HasValue ? RandomHelper.CreateSeededRandom(seed.Value) : RandomGenerator;
+        var rng = CreateInferenceRng(seed);
         var noise = SampleNoiseTensor(latent._shape, rng);
         latent = AddNoiseAtTimestep(latent, noise, startTimestep);
 

src/Diffusion/Audio/AudioLDMModel.cs

@@ -399,7 +399,7 @@ public virtual Tensor<T> TransformAudio(
         var startTimestep = Scheduler.Timesteps.Skip(startStep).FirstOrDefault();
 
         // Add noise at starting timestep
-        var rng = seed.HasValue ? RandomHelper.CreateSeededRandom(seed.Value) : RandomGenerator;
+        var rng = CreateInferenceRng(seed);
         var noise = SampleNoiseTensor(latent._shape, rng);
         latent = AddNoiseAtTimestep(latent, noise, startTimestep);
 

src/Diffusion/Audio/DiffWaveModel.cs

@@ -262,7 +262,7 @@ public virtual Tensor<T> GenerateFromMelSpectrogram(
         var shape = new[] { 1, length };
 
         // Initialize with noise
-        var rng = seed.HasValue ? RandomHelper.CreateSeededRandom(seed.Value) : RandomGenerator;
+        var rng = CreateInferenceRng(seed);
         var audio = SampleNoise(shape, rng);
 
         // Set up scheduler
@@ -300,7 +300,7 @@ public virtual Tensor<T> GenerateBatch(
     {
         var shape = new[] { batchSize, sampleLength };
 
-        var rng = seed.HasValue ? RandomHelper.CreateSeededRandom(seed.Value) : RandomGenerator;
+        var rng = CreateInferenceRng(seed);
         var audio = SampleNoise(shape, rng);
 
         Scheduler.SetTimesteps(numInferenceSteps);
@@ -398,7 +398,7 @@ public override IDiffusionModel<T> Clone()
             clone._network.ResolveLayerShapesFor(_lastInputShape);
             clone._lastInputShape = (int[])_lastInputShape.Clone();
         }
-        clone.SetParameters(GetParameters());
+        if (!clone.TryShareParametersFrom(this)) clone.SetParameters(GetParameters());
         return clone;
     }
 

src/Diffusion/Audio/MusicGenModel.cs

@@ -640,7 +640,7 @@ private Tensor<T> GenerateMusicLatent(
         }
 
         // Initialize random latent
-        var rng = seed.HasValue ? RandomHelper.CreateSeededRandom(seed.Value) : RandomGenerator;
+        var rng = CreateInferenceRng(seed);
         var latent = SampleNoiseTensor(latentShape, rng);
 
         // Set up scheduler
@@ -712,7 +712,7 @@ private Tensor<T> GenerateContinuationLatent(
         }
 
         // Initialize with partial noise (more structure from prompt)
-        var rng = seed.HasValue ? RandomHelper.CreateSeededRandom(seed.Value) : RandomGenerator;
+        var rng = CreateInferenceRng(seed);
         var noise = SampleNoiseTensor(latentShape, rng);
 
         // Blend prompt latent into initial state

src/Diffusion/Audio/RiffusionModel.cs

@@ -360,7 +360,7 @@ private Tensor<T> GenerateSpectrogramInternal(
         var latentShape = new[] { 1, RIFF_LATENT_CHANNELS, latentHeight, latentWidth };
 
         // Initialize noise
-        var rng = seed.HasValue ? RandomHelper.CreateSeededRandom(seed.Value) : RandomGenerator;
+        var rng = CreateInferenceRng(seed);
         var latents = SampleNoiseTensor(latentShape, rng);
 
         // Set up scheduler
@@ -459,7 +459,7 @@ public virtual Tensor<T> InterpolateStyles(
         var latentShape = new[] { 1, RIFF_LATENT_CHANNELS, latentHeight, latentWidth };
 
         // Initialize noise
-        var rng = seed.HasValue ? RandomHelper.CreateSeededRandom(seed.Value) : RandomGenerator;
+        var rng = CreateInferenceRng(seed);
         var latents = SampleNoiseTensor(latentShape, rng);
 
         var effectiveGuidanceScale = guidanceScale ?? GuidanceScale;

src/Diffusion/AudioDiffusionModelBase.cs

@@ -172,7 +172,7 @@ public virtual Tensor<T> GenerateFromText(
         var latentShape = new[] { 1, LatentChannels, MelChannels / VAE.DownsampleFactor, latentTimeFrames };
 
         // Generate initial noise
-        var rng = seed.HasValue ? RandomHelper.CreateSeededRandom(seed.Value) : RandomGenerator;
+        var rng = CreateInferenceRng(seed);
         var latents = DiffusionNoiseHelper<T>.SampleGaussian(latentShape, rng);
 
         // Set up scheduler
@@ -257,7 +257,7 @@ public virtual Tensor<T> TextToSpeech(
         var latentShape = new[] { 1, LatentChannels, MelChannels / VAE.DownsampleFactor, latentTimeFrames };
 
         // Generate initial noise
-        var rng = seed.HasValue ? RandomHelper.CreateSeededRandom(seed.Value) : RandomGenerator;
+        var rng = CreateInferenceRng(seed);
         var latents = DiffusionNoiseHelper<T>.SampleGaussian(latentShape, rng);
 
         // Set up scheduler
@@ -331,7 +331,7 @@ public virtual Tensor<T> AudioToAudio(
         var startTimestep = Scheduler.Timesteps.Skip(startStep).First();
 
         // Add noise to latents at starting timestep
-        var rng = seed.HasValue ? RandomHelper.CreateSeededRandom(seed.Value) : RandomGenerator;
+        var rng = CreateInferenceRng(seed);
         var noise = DiffusionNoiseHelper<T>.SampleGaussian(latentShape, rng);
         var noisyLatents = Scheduler.AddNoise(latents.ToVector(), noise.ToVector(), startTimestep);
         latents = new Tensor<T>(latentShape, noisyLatents);
@@ -388,7 +388,7 @@ public virtual Tensor<T> ContinueAudio(
         var extensionShape = new[] { inputShape[0], inputShape[1], inputShape[2], extensionLatentFrames };
 
         // Generate noise for extension
-        var rng = seed.HasValue ? RandomHelper.CreateSeededRandom(seed.Value) : RandomGenerator;
+        var rng = CreateInferenceRng(seed);
         var extensionLatents = DiffusionNoiseHelper<T>.SampleGaussian(extensionShape, rng);
 
         // Get conditioning from end of input

src/Diffusion/Control/ControlARModel.cs

@@ -156,7 +156,9 @@ public override void SetParameters(Vector<T> parameters)
     public override IDiffusionModel<T> Clone()
     {
         var clone = new ControlARModel<T>(controlType: _controlType, conditioner: _conditioner, seed: RandomGenerator.Next());
-        clone.SetParameters(GetParameters());
+        // Copy-on-write: share weight tensors with the clone (O(1)-until-write) via the global helper;
+        // fall back to the eager flat copy only if the trainable-layer structure doesn't line up 1:1.
+        if (!clone.TryShareParametersFrom(this)) clone.SetParameters(GetParameters());
         return clone;
     }
 

src/Diffusion/Control/ControlNeXtModel.cs

@@ -136,7 +136,7 @@ public override void SetParameters(Vector<T> parameters)
     public override IDiffusionModel<T> Clone()
     {
         var clone = new ControlNeXtModel<T>(controlType: _controlType, conditioner: _conditioner, seed: RandomGenerator.Next());
-        clone.SetParameters(GetParameters());
+        if (!clone.TryShareParametersFrom(this)) clone.SetParameters(GetParameters());
         return clone;
     }
 

src/Diffusion/Control/ControlNetInpaintingModel.cs

@@ -157,7 +157,7 @@ public override void SetParameters(Vector<T> parameters)
     public override IDiffusionModel<T> Clone()
     {
         var clone = new ControlNetInpaintingModel<T>(controlType: _controlType, conditioner: _conditioner, seed: RandomGenerator.Next());
-        clone.SetParameters(GetParameters());
+        if (!clone.TryShareParametersFrom(this)) clone.SetParameters(GetParameters());
         return clone;
     }
 

src/Diffusion/Control/ControlNetLiteModel.cs

@@ -147,7 +147,7 @@ public override void SetParameters(Vector<T> parameters)
     public override IDiffusionModel<T> Clone()
     {
         var clone = new ControlNetLiteModel<T>(controlType: _controlType, conditioner: _conditioner, seed: RandomGenerator.Next());
-        clone.SetParameters(GetParameters());
+        if (!clone.TryShareParametersFrom(this)) clone.SetParameters(GetParameters());
         return clone;
     }
 

src/Diffusion/Control/ControlNetModel.cs

@@ -392,7 +392,7 @@ public virtual Tensor<T> GenerateWithControl(
         var latentShape = new[] { 1, CN_LATENT_CHANNELS, latentHeight, latentWidth };
 
         // Initialize noise
-        var rng = seed.HasValue ? RandomHelper.CreateSeededRandom(seed.Value) : RandomGenerator;
+        var rng = CreateInferenceRng(seed);
         var latents = SampleNoiseTensor(latentShape, rng);
 
         // Set up scheduler
@@ -552,7 +552,7 @@ private Tensor<T> GenerateWithControlFeatures(
         var latentShape = new[] { 1, CN_LATENT_CHANNELS, latentHeight, latentWidth };
 
         // Initialize noise
-        var rng = seed.HasValue ? RandomHelper.CreateSeededRandom(seed.Value) : RandomGenerator;
+        var rng = CreateInferenceRng(seed);
         var latents = SampleNoiseTensor(latentShape, rng);
 
         // Set up scheduler
@@ -698,7 +698,7 @@ public override IDiffusionModel<T> Clone()
             clone.GetOrCreateEncoder(controlType);
         }
 
-        clone.SetParameters(GetParameters());
+        if (!clone.TryShareParametersFrom(this)) clone.SetParameters(GetParameters());
         clone.ConditioningStrength = _conditioningStrength;
 
         return clone;

src/Diffusion/Control/ControlNetPlusPlusFluxModel.cs

@@ -137,7 +137,9 @@ public override IDiffusionModel<T> Clone()
     {
         var clone = new ControlNetPlusPlusFluxModel<T>(
             controlType: _controlType, conditioner: _conditioner, rewardWeight: _rewardWeight, seed: RandomGenerator.Next());
-        clone.SetParameters(GetParameters());
+        // #1624: O(1)-until-write copy-on-write parameter share (avoids the full-model flatten copy that
+        // OOMs the 16 GB runner). Falls back to the flat copy if the structure doesn't match 1:1.
+        if (!clone.TryShareParametersFrom(this)) clone.SetParameters(GetParameters());
         return clone;
     }
 

src/Diffusion/Control/ControlNetPlusPlusModel.cs

@@ -185,7 +185,7 @@ public override IDiffusionModel<T> Clone()
             conditioner: _conditioner,
             rewardWeight: _rewardWeight,
             seed: RandomGenerator.Next());
-        clone.SetParameters(GetParameters());
+        if (!clone.TryShareParametersFrom(this)) clone.SetParameters(GetParameters());
         return clone;
     }
 

src/Diffusion/Control/ControlNetSD3Model.cs

@@ -169,7 +169,7 @@ public override IDiffusionModel<T> Clone()
             controlType: _controlType,
             conditioner: _conditioner,
             seed: RandomGenerator.Next());
-        clone.SetParameters(GetParameters());
+        if (!clone.TryShareParametersFrom(this)) clone.SetParameters(GetParameters());
         return clone;
     }
 

src/Diffusion/Control/ControlNetTileModel.cs

@@ -142,7 +142,7 @@ public override void SetParameters(Vector<T> parameters)
     public override IDiffusionModel<T> Clone()
     {
         var clone = new ControlNetTileModel<T>(conditioner: _conditioner, seed: RandomGenerator.Next());
-        clone.SetParameters(GetParameters());
+        if (!clone.TryShareParametersFrom(this)) clone.SetParameters(GetParameters());
         return clone;
     }
 

src/Diffusion/Control/ControlNetUnionProModel.cs

@@ -197,7 +197,7 @@ public override IDiffusionModel<T> Clone()
             conditioner: _conditioner,
             supportedTypes: _supportedTypes,
             seed: RandomGenerator.Next());
-        clone.SetParameters(GetParameters());
+        if (!clone.TryShareParametersFrom(this)) clone.SetParameters(GetParameters());
         return clone;
     }
 

src/Diffusion/Control/IPAdapterFaceIDPlusModel.cs

@@ -174,7 +174,7 @@ private static void AddParams(List<T> allParams, Vector<T> p)
     public override IDiffusionModel<T> Clone()
     {
         var clone = new IPAdapterFaceIDPlusModel<T>(conditioner: _conditioner, faceIdScale: _faceIdScale, seed: RandomGenerator.Next());
-        clone.SetParameters(GetParameters());
+        if (!clone.TryShareParametersFrom(this)) clone.SetParameters(GetParameters());
         return clone;
     }
 

src/Diffusion/Control/IPAdapterModel.cs

@@ -427,7 +427,7 @@ private Tensor<T> GenerateWithEmbedding(
         var latentShape = new[] { 1, IPA_LATENT_CHANNELS, latentHeight, latentWidth };
 
         // Initialize noise
-        var rng = seed.HasValue ? RandomHelper.CreateSeededRandom(seed.Value) : RandomGenerator;
+        var rng = CreateInferenceRng(seed);
         var latents = SampleNoiseTensor(latentShape, rng);
 
         // Set up scheduler
@@ -567,7 +567,7 @@ public override IDiffusionModel<T> Clone()
             conditioner: _conditioner,
             seed: RandomGenerator.Next());
 
-        clone.SetParameters(GetParameters());
+        if (!clone.TryShareParametersFrom(this)) clone.SetParameters(GetParameters());
         clone.ImagePromptWeight = _imagePromptWeight;
 
         return clone;

src/Diffusion/Control/IPAdapterPlusModel.cs

@@ -174,7 +174,7 @@ public override IDiffusionModel<T> Clone()
             conditioner: _conditioner,
             ipAdapterScale: _ipAdapterScale,
             seed: RandomGenerator.Next());
-        clone.SetParameters(GetParameters());
+        if (!clone.TryShareParametersFrom(this)) clone.SetParameters(GetParameters());
         return clone;
     }
 

src/Diffusion/Control/ReferenceOnlyModel.cs

@@ -153,7 +153,7 @@ public override void SetParameters(Vector<T> parameters)
     public override IDiffusionModel<T> Clone()
     {
         var clone = new ReferenceOnlyModel<T>(conditioner: _conditioner, referenceWeight: _referenceWeight, seed: RandomGenerator.Next());
-        clone.SetParameters(GetParameters());
+        if (!clone.TryShareParametersFrom(this)) clone.SetParameters(GetParameters());
         return clone;
     }