qualia_core_db/inference/lab/
audit_path.rs

1//! Hot-path wiring auditor — makes unfinished integration visible.
2
3use crate::inference_modes::{
4    active_inference_mode, post_turn_verify_enabled, prefer_tensor_core_gemm,
5    quant_graph_grounding_enabled, sentinel_mid_decode_enabled,
6};
7use crate::llm_bench::{
8    attention_o_fuse_enabled, attention_preproject_enabled, coop_gemv_enabled, ffn_fusion_enabled,
9    ffn_fusion_in_resident, kv_int8_enabled, resident_decode_enabled, resident_prefill_enabled,
10    resident_weights_enabled,
11};
12
13#[derive(Debug, Clone)]
14pub struct HotPathAudit {
15    pub resident_decode: bool,
16    pub resident_prefill: bool,
17    pub resident_weights: bool,
18    pub coop_gemv: bool,
19    pub ffn_fusion_flag: bool,
20    pub ffn_fusion_in_resident_decode: bool,
21    pub kv_int8: bool,
22    pub attention_preproject: bool,
23    pub attention_o_fuse: bool,
24    pub prefer_cuda_gemm: bool,
25    pub cuda_caps: bool,
26    pub mode: String,
27    pub post_turn_verify: bool,
28    pub sentinel_mid: bool,
29    pub quant_graph: bool,
30    pub timestamps_supported: bool,
31    pub gpu_profile_env: bool,
32    pub notes: Vec<String>,
33}
34
35/// Static/dynamic audit of claimed features vs what decode is configured to use.
36pub fn audit_hot_path() -> HotPathAudit {
37    let mut notes = Vec::new();
38    let ffn_flag = ffn_fusion_enabled();
39    // Set true when a resident plan was built with fused_ffn bind groups (T-A1).
40    // False until first successful plan build, or when quant is Q4_K_SOA/F16 (fallback).
41    let ffn_in_resident = ffn_fusion_in_resident();
42    if ffn_flag && !ffn_in_resident {
43        notes.push(
44            "ffn_fusion flag ON but last resident plan did not fuse (unsupported quant or plan not built yet) — T-A1 partial"
45                .into(),
46        );
47    }
48    if ffn_in_resident {
49        notes.push(
50            "T-A1/T-A1b: fused_ffn in resident (coop entry when QUALIA_LLM_COOP_GEMV on)".into(),
51        );
52    }
53
54    let cuda_caps = crate::wgsl_forge::dispatch::caps().cuda;
55    if prefer_tensor_core_gemm() && !cuda_caps {
56        notes.push("mode prefers CUDA GEMM but cuda caps=false".into());
57    }
58    if prefer_tensor_core_gemm() && cuda_caps {
59        notes.push(
60            "CUDA lane available for densify/Q4 GEMV; full hidden-on-device layer stack still open"
61                .into(),
62        );
63    }
64
65    let timestamps_supported = crate::gpu_context::shared_gpu().timestamps_supported;
66    let gpu_profile_env = std::env::var("QUALIA_LLM_GPU_PROFILE")
67        .map(|v| v == "1" || v.eq_ignore_ascii_case("true"))
68        .unwrap_or(false);
69
70    if !resident_decode_enabled() {
71        notes.push("resident_decode OFF — legacy multi-fence path".into());
72    }
73    if !coop_gemv_enabled() {
74        notes.push("coop_gemv OFF — naive GEMV".into());
75    }
76
77    HotPathAudit {
78        resident_decode: resident_decode_enabled(),
79        resident_prefill: resident_prefill_enabled(),
80        resident_weights: resident_weights_enabled(),
81        coop_gemv: coop_gemv_enabled(),
82        ffn_fusion_flag: ffn_flag,
83        ffn_fusion_in_resident_decode: ffn_in_resident,
84        kv_int8: kv_int8_enabled(),
85        attention_preproject: attention_preproject_enabled(),
86        attention_o_fuse: attention_o_fuse_enabled(),
87        prefer_cuda_gemm: prefer_tensor_core_gemm(),
88        cuda_caps,
89        mode: active_inference_mode().as_str().to_string(),
90        post_turn_verify: post_turn_verify_enabled(),
91        sentinel_mid: sentinel_mid_decode_enabled(),
92        quant_graph: quant_graph_grounding_enabled(),
93        timestamps_supported,
94        gpu_profile_env,
95        notes,
96    }
97}
98
99impl HotPathAudit {
100    pub fn format_report(&self) -> String {
101        let yn = |b: bool| if b { "yes" } else { "NO " };
102        let mut s = String::from("Hot-path wiring audit\n");
103        s.push_str(&format!("  mode:                    {}\n", self.mode));
104        s.push_str(&format!(
105            "  resident_decode:         {}\n",
106            yn(self.resident_decode)
107        ));
108        s.push_str(&format!(
109            "  resident_prefill:        {}\n",
110            yn(self.resident_prefill)
111        ));
112        s.push_str(&format!(
113            "  resident_weights:        {}\n",
114            yn(self.resident_weights)
115        ));
116        s.push_str(&format!(
117            "  coop_gemv:               {}\n",
118            yn(self.coop_gemv)
119        ));
120        s.push_str(&format!(
121            "  ffn_fusion flag:         {}\n",
122            yn(self.ffn_fusion_flag)
123        ));
124        s.push_str(&format!(
125            "  ffn_fusion in resident:  {}  (must be yes for T-A1 done)\n",
126            yn(self.ffn_fusion_in_resident_decode)
127        ));
128        s.push_str(&format!(
129            "  kv_int8:                 {}\n",
130            yn(self.kv_int8)
131        ));
132        s.push_str(&format!(
133            "  attention_preproject:    {}\n",
134            yn(self.attention_preproject)
135        ));
136        s.push_str(&format!(
137            "  attention_o_fuse:        {}\n",
138            yn(self.attention_o_fuse)
139        ));
140        s.push_str(&format!(
141            "  prefer_cuda_gemm:        {}\n",
142            yn(self.prefer_cuda_gemm)
143        ));
144        s.push_str(&format!(
145            "  cuda_caps:               {}\n",
146            yn(self.cuda_caps)
147        ));
148        s.push_str(&format!(
149            "  post_turn_verify:        {}\n",
150            yn(self.post_turn_verify)
151        ));
152        s.push_str(&format!(
153            "  sentinel_mid:            {}\n",
154            yn(self.sentinel_mid)
155        ));
156        s.push_str(&format!(
157            "  quant_graph:             {}\n",
158            yn(self.quant_graph)
159        ));
160        s.push_str(&format!(
161            "  timestamps_supported:    {}\n",
162            yn(self.timestamps_supported)
163        ));
164        s.push_str(&format!(
165            "  QUALIA_LLM_GPU_PROFILE:  {}\n",
166            yn(self.gpu_profile_env)
167        ));
168        if !self.notes.is_empty() {
169            s.push_str("  notes:\n");
170            for n in &self.notes {
171                s.push_str(&format!("    - {n}\n"));
172            }
173        }
174        s
175    }
176}
177
178#[cfg(test)]
179mod tests {
180    use super::*;
181
182    #[test]
183    fn audit_runs() {
184        let a = audit_hot_path();
185        let r = a.format_report();
186        assert!(r.contains("resident_decode"));
187        assert!(r.contains("ffn_fusion in resident"));
188    }
189}
qualia_core_db/inference/lab/audit_path.rs

qualia_core_db/inference/lab/
audit_path.rs
