Quantization Workflow Diagrams

flowchart TD S1["Step 1: Train FP32 Model\nUse standard training pipeline\nAchieve target accuracy in FP32"] S2["Step 2: Select Quantization Backend\ntorch.backends.quantized.engine = 'qnnpack'\n(ARM/mobile) or 'fbgemm' (x86/server)"] S3["Step 3: Fuse Operators\ntorch.quantization.fuse_modules()\nFuse Conv + BatchNorm + ReLU\n→ Reduces quantization rounding steps"] S4["Step 4: Assign QConfig\nmodel.qconfig = get_default_qconfig(backend)\nQConfig specifies:\n - activation observer (MinMax, Histogram, etc.)\n - weight observer (per-channel MinMax)"] S5["Step 5: Prepare Model\ntorch.quantization.prepare(model)\n→ Inserts observer modules before/after\n every quantizable operation"] S6["Step 6: Run Calibration Data\nRun 100–1000 representative samples\nthrough the prepared model\n→ Observers collect statistics"] S7["Step 7: Convert to INT8\ntorch.quantization.convert(model)\n→ Replaces float ops with int ops\n→ Freezes scale/zero_point from observers\n→ Removes observer modules"] S8["Step 8: Validate & Deploy\nMeasure accuracy on test set\nBenchmark latency & throughput\nCompare FP32 vs INT8 metrics"] S1 --> S2 S2 --> S3 S3 --> S4 S4 --> S5 S5 --> S6 S6 --> S7 S7 --> S8 style S1 fill:#4488ff,color:#fff style S5 fill:#ff9900,color:#000 style S7 fill:#44aa44,color:#fff style S8 fill:#884488,color:#fff

flowchart TD Q1["Step 1: Start from Pretrained FP32 Model\nUse converged FP32 checkpoint\nHigher initial accuracy → better final QAT result"] Q2["Step 2: Prepare for QAT\ntorch.quantization.prepare_qat(model)\nFuses Conv-BN-ReLU patterns\nInserts FakeQuantize nodes:\n - Acts on activations (after each layer)\n - Acts on weights (before each Conv/Linear)"] Q3["Step 3: Fine-tune with Simulated Quantization\nForward pass uses fake-quantized values:\n x_q = round(x/scale) * scale (FP32 domain)\nLower LR than original training (1e-5 to 1e-4)\nTypically 10–25% of original training epochs"] Q4["Step 4: STE Enables Weight Updates\nBackward pass gradient through FakeQuantize:\n ∂L/∂x ≈ ∂L/∂x_q (identity STE)\nWeights adapt to minimize loss\nwith quantization noise baked in"] Q5["Step 5: Convert to True INT8\nmodel.eval()\ntorch.quantization.convert(model)\nRemoves FakeQuantize nodes\nConverts float ops to integer ops\nFreezes learned scale/zero_point values"] Q6["Step 6: Validate & Deploy\nExpected: 0.1–0.5% accuracy drop vs FP32\nVs PTQ: 0.5–2% accuracy drop\nSame inference speed as PTQ INT8"] Q1 --> Q2 Q2 --> Q3 Q3 --> Q4 Q4 --> Q3 Q4 --> Q5 Q5 --> Q6 style Q1 fill:#4488ff,color:#fff style Q2 fill:#ff9900,color:#000 style Q4 fill:#ffdd44,color:#000 style Q5 fill:#44aa44,color:#fff style Q6 fill:#884488,color:#fff

flowchart TD IN["Input: Activation Tensor\nfrom calibration data (N samples)"] COLLECT["Collect Statistics\nPass calibration batches through model\nObservers record activation values"] HIST["Build Histogram\nBin activations into kl_bins buckets\n(captures full distribution shape)"] subgraph METHODS ["Calibration Method Choices"] direction LR MM["Min-Max\nrange = [min, max]\nFast, exact, outlier-sensitive"] PCT["Percentile\nrange = [p-th, (100-p)-th]\nClip outliers at chosen %"] MSE_C["MSE Search\nFor α in [0.8, 1.0]:\n range = [−α×max, +α×max]\n Pick α minimizing MSE"] KL_C["KL-Divergence\nFor threshold T in [max/2, max]:\n Clip to T, quantize to 256 bins\n Minimize KL(FP32 || INT8)"] end SCALE["Compute scale & zero_point\nsymmetric: scale = max_abs / 127\nasymmetric: scale = range / 255"] VALIDATE["Validate Range\nCompute SQNR for chosen range\nIf SQNR < 30 dB: try different method"] ASSIGN["Assign to Model\nFreeze scale/zero_point per layer\nReady for torch.quantization.convert()"] IN --> COLLECT COLLECT --> HIST HIST --> METHODS METHODS --> SCALE SCALE --> VALIDATE VALIDATE --> ASSIGN style METHODS fill:#f8f8f8 style MM fill:#ddeeFF style PCT fill:#ddffd4 style MSE_C fill:#fff4dd style KL_C fill:#ffd4d4

flowchart TD START["FP32 Baseline\nA_fp32 = evaluate(fp32_model, test_data)"] LAYERS["Enumerate Quantizable Layers\nL = {conv1, conv2, fc1, fc2, ...}"] FOR_EACH["For each layer L_i in L"] CLONE["Clone FP32 model\nmodel_i = deepcopy(fp32_model)"] QUANT_LAYER["Quantize ONLY layer L_i\nmodel_i.L_i.qconfig = default_qconfig\nAll other layers: qconfig = None\ntorch.quantization.prepare() → calibrate → convert()"] EVAL["Evaluate model_i\nA_i = evaluate(model_i, test_data)"] SENSITIVITY["Record sensitivity\nsens(L_i) = A_fp32 - A_i"] ALL_DONE{"All layers\nanalyzed?"} RANK["Rank layers by sensitivity\nHighest drop = Most sensitive"] THRESHOLD["Apply threshold from config:\nsensitive_layer_threshold = 0.01\n(1% accuracy drop)"] ASSIGN_FP16["sens(L_i) > threshold\n→ Assign FP16\n(2 bytes/weight)"] ASSIGN_INT8["sens(L_i) ≤ threshold\n→ Assign INT8\n(1 byte/weight)"] MIXED["Mixed Precision Model\n- Sensitive layers: FP16 compute\n- Insensitive layers: INT8 compute\n- Estimate size: Σ bytes per layer"] COMPARE["Compare vs Full INT8:\n- Size reduction achieved\n- Accuracy recovered\n- Latency profile"] START --> LAYERS LAYERS --> FOR_EACH FOR_EACH --> CLONE CLONE --> QUANT_LAYER QUANT_LAYER --> EVAL EVAL --> SENSITIVITY SENSITIVITY --> ALL_DONE ALL_DONE -->|No| FOR_EACH ALL_DONE -->|Yes| RANK RANK --> THRESHOLD THRESHOLD --> ASSIGN_FP16 THRESHOLD --> ASSIGN_INT8 ASSIGN_FP16 --> MIXED ASSIGN_INT8 --> MIXED MIXED --> COMPARE style START fill:#4488ff,color:#fff style ASSIGN_FP16 fill:#ffaa44,color:#000 style ASSIGN_INT8 fill:#ff9900,color:#000 style MIXED fill:#44aa44,color:#fff style COMPARE fill:#884488,color:#fff

flowchart LR subgraph BEFORE ["Before Folding"] direction TB C1["Conv2d\ny = W*x + b"] BN1["BatchNorm2d\ny = γ*(x-μ)/σ + β"] C1 --> BN1 end FOLD["Fold BN into Conv\nW_new = W * γ/σ\nb_new = (b-μ)*γ/σ + β"] subgraph AFTER ["After Folding"] direction TB C2["Conv2d (with bias)\ny = W_new*x + b_new"] NOTE["BN layer eliminated!\nSame mathematical output\nFewer quantized operations"] C2 --> NOTE end BEFORE --> FOLD --> AFTER style BEFORE fill:#ffe8e8 style AFTER fill:#e8ffe8