[amd-callumm]

Overview

(Work in progress)
Adds ggml_cuda_roofline_reset() call at the end of load_model(), after all initialization completes (e.g. model load, speculative decode setup, mmproj load) but before any user requests. This excludes warmup/initialization overhead from TTFT measurements to give a more representative run-to-run value, useful for benchmarking purposes.

Additional information

Following the pattern from llama-mtmd-cli (ref: PR #42) and llama-bench:

• Add roofline header include with #ifdef GGML_HIP_ROOFLINE guard
• Call ggml_cuda_roofline_reset() after params propagation
• Update CMakeLists.txt with compile definitions and include dirs

Testing

Tested on a machine equipped with a gfx1151 AMDGPU:

• Created 2 source builds of llama-server with -DGGML_HIP=ON -DGGML_HIP_ROOFLINE=ON:
  • Baseline from AMD-Ecosystem/llama.cpp:gfx11 (without these changes)
  • Second build with these changes on top of the gfx11 branch
• Using TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF, ran multiple iterations of a chat completion request with both llama-server builds
• Measured GPU time and kernel invocation count

Results of change:

• GPU invocations: 3480 --> 2774 (-20%)
• Measured GPU time: 48.7ms --> 39.5ms

NOT YET VERIFIED: TTFT or timings.prompt_ms impact.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - this fix is Claude Code-generated and human-reviewed. The agent was guided on the design based on PR mtmd-cli: reset roofline profiler after warmup prefill #42 and existing human discussion of the specific benchmarking problem.