Signed-off-by: Shijie Wang <jaywan@nvidia.com>
Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>
Signed-off-by: Shijie <jaywan@nvidia.com>
Co-authored-by: Yukun He <23156053+hyukn@users.noreply.github.com>
The performance results of some kernels could be easily affected by the warm/cold L2 cache status. To achieve more precise profiling results, the L2 cache is cleared for every execution by the circular buffer method for better benchmarking during autotuning.
Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>
* Skip the shape profile generating process if the profile has already been found in the cache under tuning mode. This is a prerequisite for nested autotuning because host overhead might be included during the profiling of the high-level op.
* Enable the profiling with CUDA graph as the default profiling method.
* Apply a heuristic method to cut off the number of repeat times of profiling according to a few-run time measurement.
Update TRTLLM Cutlass MoE kernels with ReLU2 activation.
Nemotron-6 requires ReLU2 (i.e. squared ReLU) MoE activation function.
The PR adds this and adds an API to set the activation function, in general.
The ReLU2 changes are based on this FlashInfer PR: https://github.com/flashinfer-ai/flashinfer/pull/1954.
The PR also updates the Auto Deploy MoE backend for 16-bit and FP8 from
Triton (`torch.ops.auto_deploy.triton_moe_fused`, `torch.ops.auto_deploy.triton_quant_fp8_moe`) to TRTLLM/Cutlass (`torch.ops.auto_deploy.trtllm_moe_fused`, `torch.ops.auto_deploy.trtllm_quant_fp8_moe_fused`).
Signed-off-by: Neta Zmora <96238833+nzmora-nvidia@users.noreply.github.com>
Signed-off-by: Chenghao Zhang <211069071+nvchenghaoz@users.noreply.github.com>
Co-authored-by: Chenghao Zhang <211069071+nvchenghaoz@users.noreply.github.com>
Some tunable ops require a more realistic data distribution, for instance, a shape-associated tensor. Thus, a customizable pre-hook function can be declared in the tuning config to modify the input tensor before the tuning process.
Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>
tile_tokens_dim directly depends on the num_token, which is a dynamic shape during tuning and inference. When AutoTuner prepares dummy tensors with different num_tokens, it does not update the value of tile_tokens_dim automatically. Therefore, the value stored in the AutoTuner cache is misaligned, which will introduce a lot of cache misses during inference, which hurts perf a lot.
To avoid this issue, we move the calculation of tile_tokens_dim right before kernel launching, so that the value of tile_tokens_dim is always up to date with the num_tokens of the current input tensor used for the kernel runner.
Also, the tile_tokens_dim is calculated based on the number of tokens of a tuned bucket, instead of the original token number. Because we only tune the value for the buckets, not for the raw input token number, to avoid unexpected misalignment between tile_tokens_dim and the token number.
This PR also removes the warmup requests with the extra input shapes, which are triggered in the CUDA graph warmup phase.
Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>
A redundant D2D copy is observed when enabling torch.compile for the Llama model due to the swiglu triton kernel, which brings perf overhead. Use a custom op to wrap the swiglu op to avoid this overhead.
Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>
