Skip to content

ci(TRI-1406): prepare server for 26.06 (versions, TRT 11 QA compat, enroot env fix) #8828

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
mc-nv merged 8 commits into from
Jun 11, 2026
Merged

ci(TRI-1406): prepare server for 26.06 (versions, TRT 11 QA compat, enroot env fix) #8828

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
7a8dab2
Update release versions
mc-nv Jun 9, 2026
18697f9
fix(qa): use layer.set_output_type instead of ITensor.dtype setter
mc-nv Jun 9, 2026
ba03d31
fix(qa): TRT 11 compatibility in gen_qa_identity_models.py
mc-nv Jun 9, 2026
d55d95f
fix(qa): TRT 11 compatibility across all gen_qa_*.py TRT model genera…
mc-nv Jun 9, 2026
9b47a4b
fix(qa): consistent V2/V3 dispatch in gen_qa_trt_plugin_models.py for…
mc-nv Jun 10, 2026
14b0f65
fix: Revert the ONNX Runtime version.
mc-nv Jun 10, 2026
67ecec7
fix(qa): pass TRITON_GENSRCDIR to enroot containers
mc-nv Jun 10, 2026
1a14cd4
fix(qa): document TRT 11 AttributeError pass-throughs
mc-nv Jun 10, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion TRITON_VERSION
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1 +1 @@
2.70.0dev
2.70.0
4 changes: 2 additions & 2 deletions build.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -75,8 +75,8 @@
"triton_container_version": "26.06dev",
"upstream_container_version": "26.05",
"ort_version": "1.24.4",
"ort_openvino_version": "2026.1.0",
"standalone_openvino_version": "2026.1.0",
"ort_openvino_version": "2026.2.0",
"standalone_openvino_version": "2026.2.0",
"dcgm_version": "4.5.3-1",
"rhel_py_version": "3.12.3",
}
Expand Down
22 changes: 22 additions & 0 deletions qa/common/gen_common.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -146,6 +146,28 @@ def np_to_torch_dtype(np_dtype):
return None


def trt_set_dynamic_range(tensor, lo, hi):
"""Set ITensor.dynamic_range on TRT versions that support it.

Removed in TensorRT 11+ (strongly-typed networks). Silently skip on
versions where the attribute is gone so the QA model-gen scripts stay
compatible with both old and new TRT."""
try:
tensor.dynamic_range = (lo, hi)
except AttributeError:
Comment thread
github-advanced-security[bot] marked this conversation as resolved.
Fixed
pass # ITensor.dynamic_range removed in TensorRT 11+ (strongly-typed)


def trt_cast_tensor(network, tensor, target_dtype):
"""Insert an explicit dtype cast that works on both TRT 8.5+ (add_cast)
and older TRT (add_identity + set_output_type). Returns the cast layer."""
if hasattr(network, "add_cast"):
return network.add_cast(tensor, target_dtype)
layer = network.add_identity(tensor)
layer.set_output_type(0, target_dtype)
return layer


def openvino_save_model(model_version_dir, model):
import openvino as ov

Expand Down
51 changes: 36 additions & 15 deletions qa/common/gen_qa_dyna_sequence_implicit_models.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,7 +30,12 @@
import os

import numpy as np
from gen_common import np_to_model_dtype, np_to_onnx_dtype, np_to_trt_dtype
from gen_common import (
np_to_model_dtype,
np_to_onnx_dtype,
np_to_trt_dtype,
trt_set_dynamic_range,
)

FLAGS = None
np_dtype_string = np.dtype(object)
Expand Down Expand Up @@ -388,7 +393,10 @@ def create_plan_modelfile(models_dir, model_version, max_batch, dtype, shape):
not_start = network.add_elementwise(
constant_1.get_output(0), start0, trt.ElementWiseOperation.SUB
)
not_start.set_output_type(0, trt_dtype)
# set_output_type was removed from all layers in TensorRT 11; the
# elementwise output already has trt_dtype, so this was a no-op.
if hasattr(not_start, "set_output_type"):
not_start.set_output_type(0, trt_dtype)

input_state_cond_temp = network.add_elementwise(
ready0, not_start.get_output(0), trt.ElementWiseOperation.SUM
Expand Down Expand Up @@ -527,7 +535,10 @@ def create_plan_rf_modelfile(models_dir, model_version, max_batch, dtype, shape)
not_start = network.add_elementwise(
constant_1.get_output(0), start0, trt.ElementWiseOperation.SUB
)
not_start.set_output_type(0, trt_dtype)
# set_output_type was removed from all layers in TensorRT 11; the
# elementwise output already has trt_dtype, so this was a no-op.
if hasattr(not_start, "set_output_type"):
not_start.set_output_type(0, trt_dtype)

input_state_cond_temp = network.add_elementwise(
ready0, not_start.get_output(0), trt.ElementWiseOperation.SUM
Expand All @@ -552,11 +563,19 @@ def create_plan_rf_modelfile(models_dir, model_version, max_batch, dtype, shape)

out0.get_output(0).name = "OUTPUT"
network.mark_output(out0.get_output(0))
out0.get_output(0).dtype = trt_dtype
# ITensor.dtype setter removed in TRT 11; elementwise output already has
# trt_dtype.
try:
out0.get_output(0).dtype = trt_dtype
except AttributeError:
Comment thread
github-advanced-security[bot] marked this conversation as resolved.
Fixed
pass # ITensor.dtype setter removed in TensorRT 11+

out0_state.get_output(0).name = "OUTPUT_STATE"
network.mark_output(out0_state.get_output(0))
out0_state.get_output(0).dtype = trt_dtype
try:
out0_state.get_output(0).dtype = trt_dtype
except AttributeError:
Comment thread
github-advanced-security[bot] marked this conversation as resolved.
Fixed
pass # ITensor.dtype setter removed in TensorRT 11+

in0.allowed_formats = 1 << int(trt_memory_format)
in_state0.allowed_formats = 1 << int(trt_memory_format)
Expand All @@ -566,21 +585,23 @@ def create_plan_rf_modelfile(models_dir, model_version, max_batch, dtype, shape)
out0_state.get_output(0).allowed_formats = 1 << int(trt_memory_format)

if trt_dtype == trt.int8:
in0.dynamic_range = (-128.0, 127.0)
in_state0.dynamic_range = (-128.0, 127.0)
out0.dynamic_range = (-128.0, 127.0)
out0_state.dynamic_range = (-128.0, 127.0)
start0.dynamic_range = (-128.0, 127.0)
ready0.dynamic_range = (-128.0, 127.0)

trt_set_dynamic_range(in0, -128.0, 127.0)
trt_set_dynamic_range(in_state0, -128.0, 127.0)
trt_set_dynamic_range(out0, -128.0, 127.0)
trt_set_dynamic_range(out0_state, -128.0, 127.0)
trt_set_dynamic_range(start0, -128.0, 127.0)
trt_set_dynamic_range(ready0, -128.0, 127.0)
flags = 1 << int(trt.BuilderFlag.DIRECT_IO)
flags |= 1 << int(trt.BuilderFlag.PREFER_PRECISION_CONSTRAINTS)
# TensorRT 11 removed PREFER_PRECISION_CONSTRAINTS / INT8 / FP16
# BuilderFlags (strongly-typed networks). Older TRT still has them.
if hasattr(trt.BuilderFlag, "PREFER_PRECISION_CONSTRAINTS"):
flags |= 1 << int(trt.BuilderFlag.PREFER_PRECISION_CONSTRAINTS)
if hasattr(trt.BuilderFlag, "REJECT_EMPTY_ALGORITHMS"):
flags |= 1 << int(trt.BuilderFlag.REJECT_EMPTY_ALGORITHMS)

if trt_dtype == trt.int8:
if trt_dtype == trt.int8 and hasattr(trt.BuilderFlag, "INT8"):
flags |= 1 << int(trt.BuilderFlag.INT8)
elif trt_dtype == trt.float16:
elif trt_dtype == trt.float16 and hasattr(trt.BuilderFlag, "FP16"):
flags |= 1 << int(trt.BuilderFlag.FP16)

config = builder.create_builder_config()
Expand Down
67 changes: 44 additions & 23 deletions qa/common/gen_qa_dyna_sequence_models.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,6 +36,7 @@
np_to_torch_dtype,
np_to_trt_dtype,
openvino_save_model,
trt_set_dynamic_range,
)

FLAGS = None
Expand Down Expand Up @@ -101,15 +102,26 @@ def create_plan_shape_tensor_modelfile(
resized_out0 = resize_layer.get_output(0)

shape_out0.get_output(0).name = "SHAPE_OUTPUT"
shape_out0.get_output(0).dtype = trt.int64
# ITensor.dtype setter removed in TRT 11; shape/elementwise/resize
# outputs already have the correct dtype.
try:
shape_out0.get_output(0).dtype = trt.int64
except AttributeError:
Comment thread
github-advanced-security[bot] marked this conversation as resolved.
Fixed
pass # ITensor.dtype setter removed in TensorRT 11+
network.mark_output_for_shapes(shape_out0.get_output(0))

out0.name = "OUTPUT"
out0.dtype = trt.int32
try:
out0.dtype = trt.int32
except AttributeError:
Comment thread
github-advanced-security[bot] marked this conversation as resolved.
Fixed
pass # ITensor.dtype setter removed in TensorRT 11+
network.mark_output(out0)

resized_out0.name = "RESIZED_OUTPUT"
resized_out0.dtype = trt_dtype
try:
resized_out0.dtype = trt_dtype
except AttributeError:
Comment thread
github-advanced-security[bot] marked this conversation as resolved.
Fixed
pass # ITensor.dtype setter removed in TensorRT 11+
network.mark_output(resized_out0)

shape_in0.allowed_formats = 1 << int(trt_memory_format)
Expand All @@ -121,20 +133,22 @@ def create_plan_shape_tensor_modelfile(
resized_out0.allowed_formats = 1 << int(trt_memory_format)

if trt_dtype == trt.int8:
dummy_in0.dynamic_range = (-128.0, 127.0)
resized_out0.dynamic_range = (-128.0, 127.0)
start0.dynamic_range = (-128.0, 127.0)
end0.dynamic_range = (-128.0, 127.0)
ready0.dynamic_range = (-128.0, 127.0)

trt_set_dynamic_range(dummy_in0, -128.0, 127.0)
trt_set_dynamic_range(resized_out0, -128.0, 127.0)
trt_set_dynamic_range(start0, -128.0, 127.0)
trt_set_dynamic_range(end0, -128.0, 127.0)
trt_set_dynamic_range(ready0, -128.0, 127.0)
flags = 1 << int(trt.BuilderFlag.DIRECT_IO)
flags |= 1 << int(trt.BuilderFlag.PREFER_PRECISION_CONSTRAINTS)
# TensorRT 11 removed PREFER_PRECISION_CONSTRAINTS / INT8 / FP16
# BuilderFlags (strongly-typed networks). Older TRT still has them.
if hasattr(trt.BuilderFlag, "PREFER_PRECISION_CONSTRAINTS"):
flags |= 1 << int(trt.BuilderFlag.PREFER_PRECISION_CONSTRAINTS)
if hasattr(trt.BuilderFlag, "REJECT_EMPTY_ALGORITHMS"):
flags |= 1 << int(trt.BuilderFlag.REJECT_EMPTY_ALGORITHMS)

if trt_dtype == trt.int8:
if trt_dtype == trt.int8 and hasattr(trt.BuilderFlag, "INT8"):
flags |= 1 << int(trt.BuilderFlag.INT8)
elif trt_dtype == trt.float16:
elif trt_dtype == trt.float16 and hasattr(trt.BuilderFlag, "FP16"):
flags |= 1 << int(trt.BuilderFlag.FP16)

min_prefix = []
Expand Down Expand Up @@ -353,29 +367,36 @@ def create_plan_rf_modelfile(models_dir, model_version, max_batch, dtype, shape)
out0.get_output(0).name = "OUTPUT"
network.mark_output(out0.get_output(0))

out0.get_output(0).dtype = trt_dtype
# ITensor.dtype setter removed in TRT 11; elementwise output already has
# trt_dtype.
try:
out0.get_output(0).dtype = trt_dtype
except AttributeError:
Comment thread
github-advanced-security[bot] marked this conversation as resolved.
Fixed
pass # ITensor.dtype setter removed in TensorRT 11+

in0.allowed_formats = 1 << int(trt_memory_format)
start0.allowed_formats = 1 << int(trt_memory_format)
ready0.allowed_formats = 1 << int(trt_memory_format)
out0.get_output(0).allowed_formats = 1 << int(trt_memory_format)

if trt_dtype == trt.int8:
in0.dynamic_range = (-128.0, 127.0)
out0.dynamic_range = (-128.0, 127.0)
start0.dynamic_range = (-128.0, 127.0)
end0.dynamic_range = (-128.0, 127.0)
ready0.dynamic_range = (-128.0, 127.0)
corrid0.dynamic_range = (-128.0, 127.0)

trt_set_dynamic_range(in0, -128.0, 127.0)
trt_set_dynamic_range(out0, -128.0, 127.0)
trt_set_dynamic_range(start0, -128.0, 127.0)
trt_set_dynamic_range(end0, -128.0, 127.0)
trt_set_dynamic_range(ready0, -128.0, 127.0)
trt_set_dynamic_range(corrid0, -128.0, 127.0)
flags = 1 << int(trt.BuilderFlag.DIRECT_IO)
flags |= 1 << int(trt.BuilderFlag.PREFER_PRECISION_CONSTRAINTS)
# TensorRT 11 removed PREFER_PRECISION_CONSTRAINTS / INT8 / FP16
# BuilderFlags (strongly-typed networks). Older TRT still has them.
if hasattr(trt.BuilderFlag, "PREFER_PRECISION_CONSTRAINTS"):
flags |= 1 << int(trt.BuilderFlag.PREFER_PRECISION_CONSTRAINTS)
if hasattr(trt.BuilderFlag, "REJECT_EMPTY_ALGORITHMS"):
flags |= 1 << int(trt.BuilderFlag.REJECT_EMPTY_ALGORITHMS)

if trt_dtype == trt.int8:
if trt_dtype == trt.int8 and hasattr(trt.BuilderFlag, "INT8"):
flags |= 1 << int(trt.BuilderFlag.INT8)
elif trt_dtype == trt.float16:
elif trt_dtype == trt.float16 and hasattr(trt.BuilderFlag, "FP16"):
flags |= 1 << int(trt.BuilderFlag.FP16)

min_shape = []
Expand Down
49 changes: 34 additions & 15 deletions qa/common/gen_qa_identity_models.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,6 +37,7 @@
np_to_onnx_dtype,
np_to_trt_dtype,
openvino_save_model,
trt_set_dynamic_range,
)

FLAGS = None
Expand Down Expand Up @@ -552,14 +553,18 @@ def create_plan_dynamic_rf_modelfile(
out_node = network.add_identity(in_node)

out_node.get_output(0).name = "OUTPUT{}".format(io_num)
out_node.get_output(0).dtype = trt_dtype
# Identity preserves input dtype; the ITensor.dtype setter was
# removed in TensorRT 11. Older TRT versions still accept it.
try:
out_node.get_output(0).dtype = trt_dtype
except AttributeError:
Comment thread
github-advanced-security[bot] marked this conversation as resolved.
Fixed
pass # ITensor.dtype setter removed in TensorRT 11+
network.mark_output(out_node.get_output(0))
out_node.get_output(0).allowed_formats = 1 << int(trt_memory_format)

if trt_dtype == trt.int8:
in_node.dynamic_range = (-128.0, 127.0)
out_node.get_output(0).dynamic_range = (-128.0, 127.0)

trt_set_dynamic_range(in_node, -128.0, 127.0)
trt_set_dynamic_range(out_node.get_output(0), -128.0, 127.0)
min_shape = []
opt_shape = []
max_shape = []
Expand All @@ -583,14 +588,17 @@ def create_plan_dynamic_rf_modelfile(
profile.set_shape("INPUT{}".format(io_num), min_shape, opt_shape, max_shape)

flags = 1 << int(trt.BuilderFlag.DIRECT_IO)
flags |= 1 << int(trt.BuilderFlag.PREFER_PRECISION_CONSTRAINTS)
# TensorRT 11 removed PREFER_PRECISION_CONSTRAINTS / INT8 / FP16
# BuilderFlags (strongly-typed networks). Older TRT still has them.
if hasattr(trt.BuilderFlag, "PREFER_PRECISION_CONSTRAINTS"):
flags |= 1 << int(trt.BuilderFlag.PREFER_PRECISION_CONSTRAINTS)
if hasattr(trt.BuilderFlag, "REJECT_EMPTY_ALGORITHMS"):
flags |= 1 << int(trt.BuilderFlag.REJECT_EMPTY_ALGORITHMS)
datatype_set = set([trt_dtype])
for dt in datatype_set:
if dt == trt.int8:
if dt == trt.int8 and hasattr(trt.BuilderFlag, "INT8"):
flags |= 1 << int(trt.BuilderFlag.INT8)
elif dt == trt.float16:
elif dt == trt.float16 and hasattr(trt.BuilderFlag, "FP16"):
flags |= 1 << int(trt.BuilderFlag.FP16)
config = builder.create_builder_config()
config.flags = flags
Expand Down Expand Up @@ -668,18 +676,26 @@ def create_plan_shape_tensor_modelfile(

dummy_out_node.name = "DUMMY_OUTPUT{}".format(io_num)

dummy_out_node.dtype = trt_dtype
# The ITensor.dtype setter was removed in TensorRT 11; resize and
# shape layers already produce the correct dtype, so suppress the
# AttributeError instead of changing the older-TRT behavior.
try:
dummy_out_node.dtype = trt_dtype
except AttributeError:
Comment thread
github-advanced-security[bot] marked this conversation as resolved.
Fixed
pass # ITensor.dtype setter removed in TensorRT 11+
network.mark_output(dummy_out_node)
dummy_out_node.allowed_formats = 1 << int(trt_memory_format)

out_node.get_output(0).dtype = trt.int64
try:
out_node.get_output(0).dtype = trt.int64
except AttributeError:
Comment thread
github-advanced-security[bot] marked this conversation as resolved.
Fixed
pass # ITensor.dtype setter removed in TensorRT 11+
network.mark_output_for_shapes(out_node.get_output(0))
out_node.get_output(0).allowed_formats = 1 << int(trt_memory_format)

if trt_dtype == trt.int8:
in_node.dynamic_range = (-128.0, 127.0)
out_node.get_output(0).dynamic_range = (-128.0, 127.0)

trt_set_dynamic_range(in_node, -128.0, 127.0)
trt_set_dynamic_range(out_node.get_output(0), -128.0, 127.0)
config = builder.create_builder_config()

min_prefix = []
Expand Down Expand Up @@ -707,14 +723,17 @@ def create_plan_shape_tensor_modelfile(
config.add_optimization_profile(profile)

flags = 1 << int(trt.BuilderFlag.DIRECT_IO)
flags |= 1 << int(trt.BuilderFlag.PREFER_PRECISION_CONSTRAINTS)
# TensorRT 11 removed PREFER_PRECISION_CONSTRAINTS / INT8 / FP16
# BuilderFlags (strongly-typed networks). Older TRT still has them.
if hasattr(trt.BuilderFlag, "PREFER_PRECISION_CONSTRAINTS"):
flags |= 1 << int(trt.BuilderFlag.PREFER_PRECISION_CONSTRAINTS)
if hasattr(trt.BuilderFlag, "REJECT_EMPTY_ALGORITHMS"):
flags |= 1 << int(trt.BuilderFlag.REJECT_EMPTY_ALGORITHMS)
datatype_set = set([trt_dtype])
for dt in datatype_set:
if dt == trt.int8:
if dt == trt.int8 and hasattr(trt.BuilderFlag, "INT8"):
flags |= 1 << int(trt.BuilderFlag.INT8)
elif dt == trt.float16:
elif dt == trt.float16 and hasattr(trt.BuilderFlag, "FP16"):
flags |= 1 << int(trt.BuilderFlag.FP16)
config.flags = flags

Expand Down
Loading
Loading