ggml : process data in smaller chunks in CUDA ggml_top_k() implementation to reduce temporary buffers memory usage

[fairydreaming]

Overview

This PR reduces temporary buffers memory usage in CUDA backend ggml_top_k() CUB implementation by processing input data in smaller chunks. Without this PR temporary buffers memory usage is 3 * input buffer size, allocated here:

llama.cpp/ggml/src/ggml-cuda/top-k.cu

Line 79 in d5376cf

ggml_cuda_pool_alloc<int> temp_dst_alloc(pool, ncols * nrows);

and here:

llama.cpp/ggml/src/ggml-cuda/argsort.cu

Lines 38 to 39 in d5376cf

	ggml_cuda_pool_alloc<int> temp_indices_alloc(pool, ncols * nrows);
	ggml_cuda_pool_alloc<float> temp_keys_alloc(pool, ncols * nrows);

With this PR memory usage for temporary buffers is only 3*min(input buffer size, 64MiB).

It also partially mitigates the problem of integer overflow in ncols * nrows product by lowering the amount of rows processed at once.

Fixes #24718

Additional information

For example when running this test (not present originally, I added it):

./bin/test-backend-ops -o "TOP_K(type=f32,ne=[65536,8192,1,1],k=2048,ties=0)"

without this PR memory usage in nvidia-smi goes up to 12968MiB, while with this PR it goes up only to 3048MiB.

Let's also compare the performance. Without this PR:

$ ./bin/test-backend-ops perf -o "TOP_K(type=f32,ne=[65536,8192,1,1],k=2048,ties=0)"
ggml_cuda_init: found 1 CUDA devices (Total VRAM: 97247 MiB):
  Device 0: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, compute capability 12.0, VMM: yes, VRAM: 97247 MiB
Testing 2 devices

Backend 1/2: CUDA0
  Device description: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition
  Device memory: 97247 MB (96640 MB free)

ggml_backend_cuda_graph_compute: CUDA graph warmup complete
  TOP_K(type=f32,ne=[65536,8192,1,1],k=2048,ties=0):                      32 runs - 42707.91 us/run -  2162688 kB/run -   48.29 GB/s
  Backend CUDA0: OK
Backend 2/2: CPU
  Skipping CPU backend
2/2 backends passed
OK

with this PR:

$ ./bin/test-backend-ops perf -o "TOP_K(type=f32,ne=[65536,8192,1,1],k=2048,ties=0)"
ggml_cuda_init: found 1 CUDA devices (Total VRAM: 97247 MiB):
  Device 0: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, compute capability 12.0, VMM: yes, VRAM: 97247 MiB
Testing 2 devices

Backend 1/2: CUDA0
  Device description: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition
  Device memory: 97247 MB (96640 MB free)

ggml_backend_cuda_graph_compute: CUDA graph warmup complete
  TOP_K(type=f32,ne=[65536,8192,1,1],k=2048,ties=0):                      32 runs - 42043.22 us/run -  2162688 kB/run -   49.06 GB/s
  Backend CUDA0: OK
Backend 2/2: CPU
  Skipping CPU backend
2/2 backends passed
OK

I ran test-backend-ops TOP-K tests and they all passed. Test failing in #24718 also passed:

  ...
  TOP_K(type=f32,ne=[176384,8192,1,1],k=2048,ties=0): OK
  TOP_K(type=f32,ne=[262144,8192,1,1],k=2048,ties=0): OK
  TOP_K(type=f32,ne=[843776,8192,1,1],k=2048,ties=0): OK

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: NO

[ggerganov]

I'm not sure how the cuda pool works exactly - just wondering if we actually need to have this allocation inside the loop and not one time before it?

[fairydreaming]

I'm not sure how the cuda pool works exactly - just wondering if we actually need to have this allocation inside the loop and not one time before it?

@ggerganov Good point, will try it out.

[fairydreaming]

@ORippler Do you envision any problems with organizing top-k processing the way I did it in this PR (added loop processing smaller input chunks, temporary buffer allocated outside the loop)?

[ORippler]

@ORippler Do you envision any problems with organizing top-k processing the way I did it in this PR (added loop processing smaller input chunks, temporary buffer allocated outside the loop)?

No general problems with doing it this way. Some remarks:

1. You are currently modifying the sort+view-based fallback path. Ideally, this looping should be used by ggml_cuda_op_argsort also so top_k and argsort have similar memory footprints in the CUDA backend
2. What CCCL/CTK versions are you on? I'd expect cub::DeviceTopK and cub::DeviceSegmentedTopK to offer a significantly lower memory footprint on NVGPUs (cub::DeviceSegmentedTopK should be available in CTK13.3/CCCL 3.3, but needs to be integrated into llama.cpp still): https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html) EDIT: this is not yet exposed, see Exposes DeviceBatchedTopK::{Min,Max}{Keys,Pairs} for non-deterministic, unordered, and small segments-only NVIDIA/cccl#9331

[fairydreaming]

@ORippler Do you envision any problems with organizing top-k processing the way I did it in this PR (added loop processing smaller input chunks, temporary buffer allocated outside the loop)?

No general problems with doing it this way. Some remarks:

1. You are currently modifying the sort+view-based fallback path. Ideally, this looping should be used by `ggml_cuda_op_argsort` also so top_k and argsort have similar memory footprints in the CUDA backend

@ORippler OK, will put the number of chunk rows calculation into a common function and apply this to argsort as well.