Merged
fix(wtsarq): default floating-point comparisons to cross_check #14960
fix(wtsarq): default floating-point comparisons to cross_check #14960
Merged
zsq_888Add on 3 days ago
Total: 1 files changed+8-16
@@ -14,7 +14,7 @@
14Framework choice (TTK three-leg golden spec): the operator ships a TensorFlow14Framework choice (TTK three-leg golden spec): the operator ships a TensorFlow
15frontend plugin (framework/wts_arq_tf_plugin.cpp maps OriginOpType "WtsARQ"),15frontend plugin (framework/wts_arq_tf_plugin.cpp maps OriginOpType "WtsARQ"),
16so the CPU golden is implemented with TensorFlow; numpy is used only for I/O16so the CPU golden is implemented with TensorFlow; numpy is used only for I/O
17-conversion and the torch third_party stays the independent GPU-leg reference.17+conversion and the torch third_party supplies the independent reference.
18 18 
19Delivery note: the operator is GE IR graph mode only. TensorFlow does NOT19Delivery note: the operator is GE IR graph mode only. TensorFlow does NOT
20register a callable `tf.compat.v1.WtsARQ` interface, so no TF e2e route is20register a callable `tf.compat.v1.WtsARQ` interface, so no TF e2e route is
@@ -33,14 +33,10 @@ Semantics (README.md formula, kernel.h and DESIGN.md §3.5/§3.6):
33 dequant: y = q * scale33 dequant: y = q * scale
34 34 
35Cast rule (float only, no integer outputs in @wts_arq_def.cpp):35Cast rule (float only, no integer outputs in @wts_arq_def.cpp):
36- - tolerance standard is mix_tolerance, i.e. two-leg comparison; TTK does36+ - float16 and float32 use cross_check by default. In Promote mode, the
37- not Promote, so the golden reproduces the NPU kernel width itself:37+ golden computes and returns its result in the dtype supplied by TTK.
38- LoadWiden fp16 -> fp32 for the whole computation, CAST_RINT (half to38+ - Outside Promote mode, the golden widens float16 inputs to float32
39- even) narrow back to the dispatched input dtype at the exit.39+ for computation and casts the result back to the input dtype.
40- - if a caller switches the judge to cross_check, TTK runs Promote and the
41- golden receives lifted arrays/output_dtypes; golden_mode==Promote is
42- therefore detected and the golden then does zero cast (no entry or exit
43- narrowing), letting TTK own the promotion.
44"""40"""
45 41 
46import numpy as np42import numpy as np
@@ -57,14 +53,10 @@ __spec__ = {
57 53 
58SCALE_EPS = 1.1920929e-0754SCALE_EPS = 1.1920929e-07
59 55 
60-# Precision standard: explicit ecosystem default. torch third_party below is the56+# The Torch third_party registered below supplies the reference for cross_check.
61-# independent reference for `--compare cross_check` (cross_check needs an
62-# XPU/torch provider; without one the run reports GOLDEN_FAILURE, so cross_check
63-# is not set as the default judge here). Keys must cover every dtype registered
64-# in op_host/wts_arq_def.cpp: DT_FLOAT16, DT_FLOAT.
65_TOL_KERNEL = {57_TOL_KERNEL = {
66- "float32": {"standard": "mix_tolerance"},58+ "float32": {"standard": "cross_check", "level": "L1"},
67- "float16": {"standard": "mix_tolerance"},59+ "float16": {"standard": "cross_check", "level": "L1"},
68}60}
69 61 
70 62