fix(qa): TRT 11 strongly-typed cast for plan_float* output dtype

[mc-nv]

What does the PR do?

Fixes the TRT 11 strongly-typed-network migration of gen_qa_models.py so QA plan_* model engines come out with the declared output dtype.

d898a0fe6 (TRI-1406) only inserted an explicit cast layer when the target output dtype was uint8. Float-to-float dtype changes were left to the now-removed ITensor.dtype setter, which silently no-ops on TRT 11.0. As a result, the generated engines silently inherited the elementwise op's natural output dtype (which equals the input dtype) instead of the declared output dtype, producing CI failures like:

unexpected datatype TYPE_FP32 for inference output 'OUTPUT1', expecting TYPE_FP16 for plan_float32_uint8_float16_0
unexpected datatype TYPE_FP16 for inference output 'OUTPUT0', expecting TYPE_FP32 for plan_float16_float32_float32_0

create_plan_dynamic_rf_modelfile and create_plan_fixed_rf_modelfile now insert trt_cast_tensor() whenever the elementwise op's output dtype differs from the declared output dtype — covers both float<->float and uint8 cases. The dead .dtype setter try/except AttributeError: pass blocks are removed.

Checklist

• PR title reflects the change and is of format <commit_type>: <Title>
• Changes are described in the pull request.
• Related issues are referenced.
• Populated github labels field
• Added test plan and verified test passes.
• Verified that the PR passes existing CI.
• Verified copyright is correct on all changed files.
• Added succinct git squash message before merging.
• All template sections are filled out.
• Optional: Additional screenshots for behavior/output changes with before/after.

Commit Type:

• build
• ci
• docs
• feat
• fix
• perf
• refactor
• revert
• style
• test

Related PRs:

• Predecessor: ci(TRI-1406): prepare server for 26.06 (versions, TRT 11 QA compat, enroot env fix) #8828 (TRI-1406, commit d898a0fe6 — original TRT 11 QA migration)

Where should the reviewer start?

qa/common/gen_qa_models.py:

• create_plan_dynamic_rf_modelfile (around line 109): cast block replaces the uint8-only cast + dead .dtype setter.
• create_plan_fixed_rf_modelfile (around line 412): cast block added, dead .dtype setter removed.

Test plan:

Reproduction (before patch) on r26.06 HEAD with tensorrt:26.06-py3-base (TRT 11.0):

TENSORRT_IMAGE=gitlab-master.nvidia.com:5005/dl/dgx/tensorrt:26.06-py3-base \
PYTORCH_IMAGE=gitlab-master.nvidia.com:5005/dl/dgx/pytorch:26.06-py3-base \
bash -xe ./qa/common/gen_qa_model_repository

# Then inspect engines via trtexec --loadEngine=<model.plan>

Engines produced before patch (mismatch against config.pbtxt):

Model	OUTPUT0 (config / engine)	OUTPUT1 (config / engine)
plan_float32_uint8_float16	UINT8 / uint8	FP16 / fp32 ✗
plan_float16_float32_float32	FP32 / fp16 ✗	FP32 / fp16 ✗

Engines produced after patch (all aligned):

Model	OUTPUT0 (config / engine)	OUTPUT1 (config / engine)
plan_float32_uint8_float16	UINT8 / uint8 ✓	FP16 / fp16 ✓
plan_nobatch_float32_uint8_float16	UINT8 / uint8 ✓	FP16 / fp16 ✓
plan_float16_float32_float32	FP32 / fp32 ✓	FP32 / fp32 ✓
plan_nobatch_float16_float32_float32	FP32 / fp32 ✓	FP32 / fp32 ✓

• CI Pipeline ID: (to be added — pipeline 54492155 originally exhibited the failure across 11 jobs; a new run from a regenerated 26.06_* QA repo is needed for end-to-end confirmation.)

Caveats:

• QA /data/inferenceserver/26.06_* mounts that CI consumes must be regenerated from this PR before the affected L0_infer* / L0_server_status jobs will go green. The fix is in the model-generation script; the published model artifacts on shared storage still hold the broken engines until republished.
• Other generator scripts (gen_qa_identity_models.py, gen_qa_sequence_models.py, gen_qa_implicit_models.py, gen_qa_dyna_sequence_models.py, gen_qa_dyna_sequence_implicit_models.py, gen_qa_reshape_models.py, gen_qa_sequence_models.py) still carry the same dead out.dtype = … / except AttributeError: pass pattern. Audit shows most are harmless because input/output dtypes match in those generators, but a follow-up sweep is recommended.

Background

TRT 11.0 introduced strongly-typed networks (TRT release notes). As part of that change:

• ITensor.dtype is no longer a writable setter — engine output dtype is inferred from the network graph.
• ILayer.set_output_type was removed on identity layers in favour of network.add_cast().
• BuilderFlag.PREFER_PRECISION_CONSTRAINTS, FP16, INT8 were removed (strongly-typed networks make them redundant).

d898a0fe6 added a trt_cast_tensor() helper to gen_common.py and used it for uint8 conversion in create_plan_dynamic_rf_modelfile, but did not cover the float→float dtype change case in either create_plan_dynamic_rf_modelfile or create_plan_fixed_rf_modelfile. This PR completes the migration for those two functions. The same audit was performed on the other QA generators — most do not change dtypes between input and output, so the dead setter is a no-op there; no functional bugs are blocking the immediate CI failures.

Observed failure footprint on pipeline 54492155 (branch 26.06-devel): 11 affected jobs covering SBSA A100, SBSA GB300, AGX-Thor, RTX50-rhel, and the base L0_server_status.

Related Issues: (use one of the action keywords Closes / Fixes / Resolves / Relates to)

• Relates to TRI-1406 (this is a follow-up fix to that work)
• TRI-1417 (this PR)