feat(quantize): AdaRound INT4 learned-rounding PTQ + dimensional Quantizer spec#44
Closed
nathanhubens wants to merge 3 commits into
Closed
feat(quantize): AdaRound INT4 learned-rounding PTQ + dimensional Quantizer spec#44nathanhubens wants to merge 3 commits into
nathanhubens wants to merge 3 commits into
Conversation
…tizer spec AdaRound (Nagel 2020) — the first validated safe-quant technique. Learns per-weight round up/down to minimize output reconstruction, so INT4 holds accuracy where round-to-nearest collapses. - adaround_quantize(model, calibration_dl, w_bits=4, granularity='channel', ...) — accepts a fastai DataLoaders, a torch DataLoader, or a list of batches. - Measured (ResNet-18): INT4+AdaRound recovers the RTN cliff (+26pt full budget -> ~1pt below INT8). Architecture-dependent: viable for redundant CNNs (ResNet-class), NOT efficient/depthwise nets (route those to INT8). INT4 = size lever, not CPU speed. - Tests: init==RTN correctness gate, reconstruction-improves, failure modes (no-calibration -> clear error, list-of-batches accepted), + a #|slow ResNet-18 recovery integration test. - Shaped as criteria='adaround' toward the dimensional Quantizer (DIMENSIONAL_DESIGN.md: granularity x context x criteria x schedule x dtype, mirroring Sparsifier).
Aligns the weight grid to the ONNX/ORT signed-INT4 convention so exported integers land on the deployment grid, and uses all 16 INT4 values (strict quality improvement). weight_scale now returns (scale, qmin, qmax); rtn_quant/_hard_quant_from_V/adaround_layer clamp to [qmin, qmax]. The init==RTN correctness gate is preserved (floor-based round-half-up so RTN matches the hard-round on the new grid). Full nbdev-test green (fast + slow). Lifts standard ONNX-export accuracy 71.75 -> 74.57 (ResNet-18). NOTE: grid alignment alone is NOT lossless (ORT recalibrates scale from the quantized weights' absmax) — the faithful QDQ-export helper (bake exact int4+scale, verified 79.11 ~ fake-quant 78.34) is the true deployment path (follow-up).
…nt path) AdaRound's learned rounding survives to a real INT4 ONNX ONLY if you write the exact int4+scale yourself — standard exporters recalibrate the per-channel scale and shed 4-7pt (measured). This closes the deployment gap: - adaround_quantize now stashes model._adaround_qparams (exact per-layer scale + qmin/qmax), captured pre-quant (the scale is unrecoverable once W <- scale*int). - adaround_to_onnx(model, sample, path): FP32 ONNX -> ORT QDQ structure -> overwrite the scale + int4 weight initializers with AdaRound's EXACT values -> a QDQ ONNX that carries the learned rounding. _pack_int4/_unpack_int4 helpers. - Verified faithful (slow test): ResNet-18 INT4, fake-quant 66.41 -> deployed ONNX 67.19 (delta -0.78pt, matches the standalone -0.76); int4 ONNX 5.99MB vs FP32 46.75MB (7.8x); dequant weight error 1.2e-7. Completes the loop: AdaRound accuracy -> real INT4 artifact at the SAME accuracy -> measurable via fasterrecipes evaluate_accuracy. nbdev-test green (fast+slow).
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The first validated safe-quant technique becomes a real fasterai feature — built into a coherent dimensional design rather than as a one-off.
AdaRound (
criteria='adaround')Learned-rounding INT4 PTQ (Nagel 2020): learns per-weight round-up/down to minimize output reconstruction, so INT4 holds accuracy where round-to-nearest collapses.
adaround_quantize(model, calibration_dl, w_bits=4, granularity='channel', act_bits=None, ...)— accepts a fastaiDataLoaders, a torchDataLoader, or a plain list of batches.Dimensional
Quantizerspec (nbs/quantize/DIMENSIONAL_DESIGN.md)North-star mirroring
Sparsifier— every quant technique is a composable value on an axis (granularity × context × criteria × schedule × dtype), not a bespoke function. AdaRound is the firstcriteria; the strandedquantize_mixedwork →context='mixed'; SmoothQuant → a pre-quant transform. Same mechanism/decision split as pruning (fasterai applies; fasterrecipes decides mixed-precision via sensitivity) → one allocation engine for both prune-ratio AND bit-width.Validation
#| slowtest green: real +14 pt ResNet-18 Imagenette recovery (short notebook budget; full budget ≈ +26 pt).granularitystring API.