[Feature]: AutoDeploy: investigate fp8 gemms that do not have pad to 16 requirement

[lucaslie]

🚀 The feature, motivation and pitch

Follow-up to #8811

Instead of padding the existing fp8 GEMM kernel we use, let's investigate alternative kernels (e.g. trtllm fp8 kernel) that can handle unpadded shapes that are not mod 16 natively

Alternatives

No response

Additional context

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and checked the documentation and examples for answers to frequently asked questions.

[lucaslie]

This is the trt-llm PT kernel for fp8:

TensorRT-LLM/tensorrt_llm/_torch/custom_ops/torch_custom_ops.py

Lines 334 to 357 in 9c4432f

	@torch.library.custom_op("trtllm::fp8_rowwise_gemm", mutates_args=())
	def fp8_rowwise_gemm(
	act: torch.Tensor,
	weight: torch.Tensor,
	act_scale: torch.Tensor,
	weight_scale: torch.Tensor,
	output_dtype: torch.dtype,
	to_userbuffers: bool = False,
	) -> torch.Tensor:
	tuner = AutoTuner.get()
	# allocate workspace for profiling
	fp8_rowwise_gemm_runner = FP8RowwiseGemmRunner(to_userbuffers, output_dtype)
	_, best_tactic = tuner.choose_one(
	"trtllm::fp8_rowwise_gemm::gemm",
	[fp8_rowwise_gemm_runner],
	FP8RowwiseGemmRunner.tuning_config,
	[act, weight, act_scale, weight_scale],
	)
	return fp8_rowwise_gemm_runner(
	inputs=[act, weight, act_scale, weight_scale], tactic=best_tactic)