Skip to main content

qualia_core_db/gguf_bridge/
native_resident_adapter.rs

1use super::*;
2use etalon_inference::native_resident::{ResidentProvider, ResidentResources};
3impl ResidentProvider for QTensorEngine {
4    type Mask = [u32; KV_ATTENTION_MASK_WORDS];
5    fn resident_resources(&self) -> ResidentResources<'_> {
6        ResidentResources {
7            add_pipeline: self.elem_gpu_pipeline(ELEM_OP_ADD_RESIDUAL),
8            attention_mask_buf: self.attention_mask_buf.as_ref(),
9            attention_pipeline: &self.attention_pipeline,
10            coop_gemv_mr_pipeline: &self.coop_gemv_mr_pipeline,
11            coop_gemv_pipeline: &self.coop_gemv_pipeline,
12            coop_gemv_residual_mr_pipeline: &self.coop_gemv_residual_mr_pipeline,
13            coop_gemv_residual_pipeline: &self.coop_gemv_residual_pipeline,
14            coop_gemv_residual_warp_pipeline: &self.coop_gemv_residual_warp_pipeline,
15            coop_gemv_warp_pipeline: &self.coop_gemv_warp_pipeline,
16            dual_gemv_mr_pipeline: &self.dual_gemv_mr_pipeline,
17            dual_gemv_pipeline: &self.dual_gemv_pipeline,
18            elem_silu_mul_pipeline: &self.elem_silu_mul_pipeline,
19            ffn_fused_coop_pipeline: &self.ffn_fused_coop_pipeline,
20            ffn_fused_mr_pipeline: &self.ffn_fused_mr_pipeline,
21            ffn_fused_pipeline: &self.ffn_fused_pipeline,
22            ffn_fused_warp_pipeline: &self.ffn_fused_warp_pipeline,
23            output_topk_pipeline: self.output_topk_pipeline.as_ref(),
24            pipeline: &self.pipeline,
25            rms_pipeline: self.elem_gpu_pipeline(ELEM_OP_RMS_NORM),
26            topk_cand_idx_buf: self.topk_cand_idx_buf.as_ref(),
27            topk_cand_val_buf: self.topk_cand_val_buf.as_ref(),
28            triple_gemv_pipeline: &self.triple_gemv_pipeline,
29        }
30    }
31    fn gpu_device(&self) -> &etalon::Device {
32        QTensorEngine::gpu_device(self)
33    }
34    fn gpu_queue(&self) -> &etalon::Queue {
35        QTensorEngine::gpu_queue(self)
36    }
37    fn submit(&self, command: etalon::CommandBuffer) -> bool {
38        self.submit_commands([command]).is_some()
39    }
40    fn poll_wait(&self) -> bool {
41        let result = self
42            .gpu_device()
43            .poll(etalon::PollType::wait_indefinitely())
44            .is_ok();
45        GPU_WAIT_COUNT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
46        result
47    }
48    fn attention_kv_mask(&self, token: u32, capacity: u32) -> (Self::Mask, u32) {
49        crate::compute_universe::attention_kv_mask_u32(token, capacity)
50    }
51    fn dual_multirow(&self) -> bool {
52        matches!(
53            super::execution_env::var("QUALIA_LLM_DUAL_MR")
54                .ok()
55                .as_deref(),
56            Some("1") | Some("true")
57        )
58    }
59    fn coop_gemv_workgroups(&self, rows: u32) -> u32 {
60        crate::llm_bench::coop_gemv_workgroups(rows)
61    }
62    fn pass_writes(&self) -> Option<etalon::ComputePassTimestampWrites<'static>> {
63        crate::llm_gpu_profiler::pass_writes_both()
64    }
65    fn resolve(&self, encoder: &mut etalon::CommandEncoder) {
66        crate::llm_gpu_profiler::resolve(encoder);
67    }
68    fn accumulate_fused(&self) {
69        crate::llm_gpu_profiler::accumulate(crate::llm_gpu_profiler::Phase::FusedBlock);
70    }
71}