Function preferred_dtype

tests/unit/common.py:512–518 · view source on GitHub ↗

()

Source from the content-addressed store, hash-verified

510
511	# bf16 > fp16 > fp32
512	def preferred_dtype():
513	if get_accelerator().is_bf16_supported():
514	return torch.bfloat16
515	elif get_accelerator().is_fp16_supported():
516	return torch.float16
517	else:
518	return torch.float32
519
520
521	class EnableDeterminism:

checkpoint_correctness_verificationFunction · 0.90

initialize_distributedFunction · 0.90

test_scale_validates_scalar_lossMethod · 0.90

testMethod · 0.90

init_tp_engineFunction · 0.90

assert_close_for_preferred_dtypeFunction · 0.90

test_custom_patterns_applied_via_configMethod · 0.90

test_use_default_specs_false_skips_unmatched_layersMethod · 0.90

test_custom_module_replacement_with_patternsMethod · 0.90

test_custom_pattern_disables_fused_qkv_heuristicMethod · 0.90

test_gate_up_fused_weight_partitionMethod · 0.90

test_gqa_uneven_qkv_fused_weight_partitionMethod · 0.90

get_acceleratorFunction · 0.90

is_bf16_supportedMethod · 0.45

is_fp16_supportedMethod · 0.45

initialize_distributedFunction · 0.72

test_scale_validates_scalar_lossMethod · 0.72

testMethod · 0.72

init_tp_engineFunction · 0.72

assert_close_for_preferred_dtypeFunction · 0.72

test_custom_patterns_applied_via_configMethod · 0.72

test_use_default_specs_false_skips_unmatched_layersMethod · 0.72

test_custom_module_replacement_with_patternsMethod · 0.72

test_custom_pattern_disables_fused_qkv_heuristicMethod · 0.72

test_gate_up_fused_weight_partitionMethod · 0.72

test_gqa_uneven_qkv_fused_weight_partitionMethod · 0.72

test_gqa_uneven_qkv_fused_forwardMethod · 0.72

searching dependent graphs…