Expand TRT decoder YAML config for composite decoding [depends on PR #524]#536
Expand TRT decoder YAML config for composite decoding [depends on PR #524]#536wsttiger wants to merge 26 commits into
Conversation
…output
Add a "predecoder" execution mode to the TensorRT decoder so it can be
chained with a second decoder (e.g. PyMatching) and return logical-frame
observables directly. The TRT model is assumed to emit a single output
that concatenates [pre_L (num_observables entries), residual_dets (rest)].
New constructor parameters:
- "batch_size": required when the ONNX model has a dynamic batch dim.
Used to size the optimization profile and pre-allocate I/O buffers.
- "global_decoder" + "global_decoder_params": optional decoder name and
params for a follow-up decoder run on the residual_dets portion of
the TRT output. Created with the same H passed to trt_decoder.
- "O": observables matrix (num_observables x block_size). Enables
decode()/decode_batch() to return the predicted logical frame.
Number of observables is inferred from O.shape()[0].
Decode behavior matrix:
- no global_decoder, no O -> raw TRT output (unchanged).
- no global_decoder, O -> return the pre_L prefix only.
- global_decoder, no O -> entire output -> global_decoder.result.
- global_decoder, O -> residual -> global_decoder; return
pre_L XOR global_decoder.logical_frame.
Constructor validation when O is set:
- output_size_per_sample >= num_observables, and
- when global_decoder_ is set,
output_size_per_sample == num_observables + global_decoder.syndrome_size.
Other changes:
- Dynamic batch support: setInputShape per call when the model's batch
dim is -1; ONNX builder now installs a min/opt/max optimization
profile when "batch_size" is provided.
- Split decode_batch into a typed decode_batch_impl<float|uint8_t> for
cleaner dtype dispatch (engine I/O dtypes float32 / uint8 unchanged).
- Better INFO logging: total non-zero input vs residual detector counts
per batch to help diagnose predecoder behavior.
Signed-off-by: Ben Howe <bhowe@nvidia.com>
Add a realtime test/demo that initializes the TensorRT decoder from an ONNX predecoder model with PyMatching configured as the global decoder. The driver loads detector, observable, parity-check, observable, and prior data from the Stim export bundle, decodes samples through the composite TRT+PyMatching path, and reports latency, throughput, correctness, and residual-syndrome diagnostics. Register the new test_trt_decoder_composite target when TensorRT, realtime, and the TRT decoder plugin are available. Signed-off-by: Scott Thornton <wsttiger@gmail.com>
Add YAML/config support for TRT decoder runtime options including batch size, CUDA graph execution, global decoder selection, and PyMatching-specific global decoder parameters. Wire realtime decoder construction so TRT configs receive the top-level observable matrix from O_sparse, and pass the same O matrix into PyMatching global decoder params for composite observable decoding. Expose the new config fields through Python bindings and heterogeneous_map round-tripping. Extend YAML tests for TRT config round-trip, runtime parameter conversion, and O_sparse-to-O injection. Update test_trt_decoder_composite to support an optional --config-yaml path, allowing the existing composite demo to construct and run a real TRT+PyMatching decoder directly from YAML while preserving the original manual CLI path. Signed-off-by: Scott Thornton <wsttiger@gmail.com>
…yaml # Conflicts: # libs/qec/unittests/realtime/CMakeLists.txt # libs/qec/unittests/realtime/test_trt_decoder_composite.cpp
Replace the TRT decoder's hardcoded optional PyMatching global decoder params with a tagged global_decoder_config variant. Preserve PyMatching as the current supported concrete config while using std::monostate for the unset case. Update heterogeneous-map conversion, YAML mapping, and Python bindings so the existing PyMatching YAML/Python surface continues to round-trip. Extend the YAML unit test to verify the PyMatching variant arm is selected and still produces the expected runtime parameter map. Signed-off-by: Scott Thornton <wsttiger@gmail.com>
…yaml # Conflicts: # libs/qec/python/bindings/py_decoding_config.cpp
Signed-off-by: Scott Thornton <wsttiger@gmail.com>
Signed-off-by: Scott Thornton <wsttiger@gmail.com>
wsttiger
force-pushed
the
update_trt_decoder_yaml
branch
from
6c2eefc to
26be6b4
Compare
| @@ -0,0 +1,91 @@ | |||
| # ============================================================================ # | |||
There was a problem hiding this comment.
This file sits at the root of the repo. I'm not sure if it's intentional? There are also several hardcoded paths used in this file. From the comment, it seems like this file is meant to be a draft. Is it ready for review or should it be removed from the PR?
Signed-off-by: Melody Ren <melodyr@nvidia.com>
The trt_decoder constructed with an "O" observable matrix projects to observables internally, so it must report decode_result_type::decode_to_obs to enqueue_syndrome(). Set the result type where decode_to_observables_ is enabled, and assert it in the composite test. Signed-off-by: Melody Ren <melodyr@nvidia.com>
Signed-off-by: Melody Ren <melodyr@nvidia.com>
A monostate global_decoder_params (no global decoder attached) was being
mutated into a default pymatching_decoder_config across a serialize ->
deserialize cycle, through two independent serialization layers:
1. heterogeneous_map: to_heterogeneous_map() emitted an empty
global_decoder_params map whenever global_decoder was set but the
params were monostate, which read back as a pymatching config.
2. YAML MappingTraits (the path used by to_yaml_str/from_yaml_str, and
thus by save_dem/load_dem): mapOptional emitted an empty
'global_decoder_params: {}' for the monostate case, which read back
into a default pymatching config.
Both layers now emit nothing for monostate. Any runtime need for an empty
params map is handled in prepare_decoder_params (realtime_decoding.cpp),
not in serialization. The heterogeneous_map path also rejects a params
map that carries global_decoder_params without a global_decoder.
Add regression tests: monostate round-trips unchanged through both YAML
and heterogeneous_map and emits no params key; params-without-decoder
throws.
Signed-off-by: Melody Ren <melodyr@nvidia.com>
Signed-off-by: Melody Ren <melodyr@nvidia.com>
Two follow-ups to the monostate round-trip fix: 1. prepare_decoder_params now synthesizes an empty global_decoder_params map for a pymatching global decoder before the O_sparse early return. Since serialization stopped emitting an empty params map for monostate, a global decoder configured to run on residual detectors without an O matrix was no longer attached by the plugin (which requires both global_decoder and global_decoder_params keys). This is a documented, valid configuration, so restore it in runtime prep where it belongs. 2. trt_decoder_config::to_heterogeneous_map now throws when global_decoder_params is set but global_decoder is not, matching the rejection already enforced by from_heterogeneous_map. Add regression tests for both. Signed-off-by: Melody Ren <melodyr@nvidia.com>
Signed-off-by: Melody Ren <melodyr@nvidia.com>
Signed-off-by: Melody Ren <melodyr@nvidia.com>
Signed-off-by: Melody Ren <melodyr@nvidia.com>
Signed-off-by: Melody Ren <melodyr@nvidia.com>
Signed-off-by: Melody Ren <melodyr@nvidia.com>
|
The following were added on top of 26be6b4:
|
Signed-off-by: Melody Ren <melodyr@nvidia.com>
Signed-off-by: Melody Ren <melodyr@nvidia.com>
|
/ok to test e157661 |
Resolve conflicts by keeping both PyMatching config structs for now: - pymatching_decoder_config (trt nested global decoder, PR536) and Vedika's pymatching_config (standalone realtime decoder, NVIDIA#614) coexist. - Kept the global_decoder_config variant + its serialization + YAML traits. - realtime_decoding.cpp: unioned includes; kept both prepare_decoder_params and Chuck's new get_realtime_session (NVIDIA#609). Follow-up: unify on pymatching_config (drop pymatching_decoder_config). Signed-off-by: Melody Ren <melodyr@nvidia.com>
Signed-off-by: Melody Ren <melodyr@nvidia.com>
|
/ok to test 33057f5 |
3 participants
Add YAML/config support for TRT decoder runtime options including batch size,
CUDA graph execution, global decoder selection, and PyMatching-specific global
decoder parameters. Wire realtime decoder construction so TRT configs receive
the top-level observable matrix from O_sparse, and pass the same O matrix into
PyMatching global decoder params for composite observable decoding.
Expose the new config fields through Python bindings and heterogeneous_map
round-tripping. Extend YAML tests for TRT config round-trip, runtime parameter
conversion, and O_sparse-to-O injection.
Update test_trt_decoder_composite to support an optional --config-yaml path,
allowing the existing composite demo to construct and run a real TRT+PyMatching
decoder directly from YAML while preserving the original manual CLI path.