[CUDA] Remove TensorRT fused causal attention kernels

[tianleiwu]

Description

This PR removes the TensorRT fused causal attention kernels (the fmha_v2_*_Causal_* and fmha_v2_flash_attention_*_Causal_* cubins) and all of the code paths that selected them from the CUDA Attention operator.

These causal fused kernels were disabled by default (since #14732) and were only reachable via the opt-in ORT_ENABLE_FUSED_CAUSAL_ATTENTION environment variable / TRT_CAUSAL_ATTENTION backend bit. They used fp16 accumulation, which can cause accuracy drops, and have been superseded by flash attention, memory-efficient attention, and cuDNN SDPA. Removing them deletes ~1.27M lines of generated cubin source and simplifies the attention dispatch logic.

Key Changes

• Removed cubin sources: Deleted all causal/fmha_v2_fp16_Causal_* and flash_attention/fmha_v2_flash_attention_fp16_Causal_* generated cubin files (70+ files).
• Dispatch simplification (attention.cc): Removed the is_unidirectional_ / causal fused-runner branch in ComputeInternal; the fused runner path now only handles the BERT (non-causal) case.
• Kernel options (attention_kernel_options.cc, attention_kernel_options.h): Removed use_trt_causal_attention_, UseTrtCausalAttention(), the TRT_CAUSAL_ATTENTION debug print, and the causal argument of SetTrtFusedKernel.
• QKV format (attention_common.h, attention_prepare_qkv.cu): Removed the Q_K_V_BNSH_QKV_BS3NH format and the fused-causal gemm-buffer-with-bias preparation path.
• Runner API (mha_runner.cu, mha_runner.h, fused_multihead_attention_v2.h): Dropped the causal parameter from FusedMHARunnerFP16v2::Create / IsSupported and removed the causal kernel metadata.
• Env var removed: ORT_ENABLE_FUSED_CAUSAL_ATTENTION (kEnableFusedCausalAttention) is no longer recognized.
• Callers updated: multihead_attention.cc, packed_attention.cc, packed_multihead_attention.cc, attention_impl.cu updated to the new no-causal signatures.
• Python helpers: Removed stale ORT_ENABLE_FUSED_CAUSAL_ATTENTION references from the transformers benchmark helper and stable diffusion benchmark.
• Tests updated: attention_op_test.cc and attention_kernel_options_test.cc no longer set/assert the causal-fused option.

Motivation and Context

The fused causal kernels were off by default, carried potential fp16-accumulation accuracy risk, and added a large amount of generated cubin source to the repo. Causal attention is already well covered by flash attention, memory-efficient attention, and cuDNN SDPA, so these kernels can be safely removed to reduce binary size and simplify maintenance.

Testing

• Build the CUDA EP and run the attention contrib op tests (ContribOpAttentionTest.*, including Causal_EmptyPastState).
• Run AttentionKernelOptionsTest.* to verify the kernel-option parsing no longer references the causal backend.

[Copilot]

⚠️ Not ready to approve

There are a couple of remaining inconsistencies/redundant test paths after causal-kernel removal that should be cleaned up to match the PR’s stated behavior and keep tests deterministic.

Pull request overview

This PR removes the TensorRT fused causal attention kernel assets and eliminates the corresponding CUDA Attention dispatch/option plumbing, consolidating causal attention execution onto existing paths (flash attention, memory-efficient attention, cuDNN SDPA, or unfused).

Changes:

• Removed TRT fused causal attention selection/metadata and simplified fused-runner logic to the non-causal (BERT) case.
• Deleted causal-only QKV formats and runner APIs (dropped causal parameters, removed causal kernel meta entries).
• Updated tests and Python benchmarking helpers to stop referencing the removed ORT_ENABLE_FUSED_CAUSAL_ATTENTION behavior.

File summaries

File	Description
onnxruntime/test/providers/cuda/test_cases/attention_kernel_options_test.cc	Drops causal-backend assertions/env usage from kernel-option tests.
onnxruntime/test/contrib_ops/attention_op_test.cc	Removes fused-causal env usage in causal Attention test scaffolding.
onnxruntime/python/tools/transformers/models/stable_diffusion/benchmark.py	Removes legacy workaround that set ORT_ENABLE_FUSED_CAUSAL_ATTENTION.
onnxruntime/python/tools/transformers/benchmark_helper.py	Removes ORT_ENABLE_FUSED_CAUSAL_ATTENTION from env var enumeration.
onnxruntime/contrib_ops/cuda/bert/tensorrt_fused_multihead_attention/mha_runner.h	Removes causal from TRT fused runner APIs/state.
onnxruntime/contrib_ops/cuda/bert/tensorrt_fused_multihead_attention/mha_runner.cu	Deletes causal runner setup and causal-mask routing; updates IsSupported/Create/Run.
onnxruntime/contrib_ops/cuda/bert/tensorrt_fused_multihead_attention/fused_multihead_attention.h	Removes unused causal-mask parameter from kernel run interface (cleanup).
onnxruntime/contrib_ops/cuda/bert/tensorrt_fused_multihead_attention/fused_multihead_attention_v2.h	Removes causal cubin externs and causal kernel meta entries; simplifies hashing/signatures.
onnxruntime/contrib_ops/cuda/bert/packed_multihead_attention.cc	Updates debug-info API calls for fused runner (no causal arg).
onnxruntime/contrib_ops/cuda/bert/packed_attention.cc	Updates TRT fused runner IsSupported/Create call sites and debug-info API.
onnxruntime/contrib_ops/cuda/bert/multihead_attention.cc	Updates TRT fused runner calls and debug-info API (no causal arg).
onnxruntime/contrib_ops/cuda/bert/attention.h	Removes member storing fused-causal enablement.
onnxruntime/contrib_ops/cuda/bert/attention.cc	Removes fused-causal selection branch; fused runner now only for non-unidirectional path.
onnxruntime/contrib_ops/cuda/bert/attention_prepare_qkv.cu	Removes fused-causal QKV format handling and bias-to-gemm-buffer path.
onnxruntime/contrib_ops/cuda/bert/attention_kernel_options.h	Removes causal option surface and updates debug-info API signature.
onnxruntime/contrib_ops/cuda/bert/attention_kernel_options.cc	Stops parsing/printing TRT causal option; updates debug-info classification.
onnxruntime/contrib_ops/cuda/bert/attention_impl.cu	Removes fused-causal execution path and tightens QKV format assertions.
onnxruntime/contrib_ops/cpu/bert/attention_common.h	Removes fused-causal env var constant and QKV format enum entry (CPU-side shared defs).

Copilot's findings

• Files reviewed: 18/88 changed files
• Comments generated: 3

Note

Your feedback helps us improve the quality of this feature.
Please use 👍 or 👎 to tell us whether this assessment is correct.

[Copilot]

⚠️ Human review recommended

It removes a broad CUDA/TRT attention dispatch surface (kernels, options, runner APIs, and tests), and warrants human verification of runtime/kernel routing across supported GPU/EP configurations.

Copilot's findings

• Files reviewed: 20/90 changed files
• Comments generated: 0 new

Note

Your feedback helps us improve the quality of this feature.
Please use 👍 or 👎 to tell us whether this assessment is correct.