metrale_model_engine/model/trait_impl/
impl_lifecycle.rs

1// SPDX-License-Identifier: MIT OR Apache-2.0
2
3//! 2026-09-26: `impl ModelLifecycle for TransformerModel`, mostly delegating to `<method>_dispatch`
4//! helpers in the sibling modules.
5//!
6//! Owner: model-engine.
7//! Invariants: the ones in `trait_impl/mod.rs`.
8
9use anyhow::Result;
10
11use crate::model::types::TransformerModel;
12use crate::traits::{ModelLifecycle, SequenceState};
13
14impl ModelLifecycle for TransformerModel {
15    fn teardown(&mut self) -> Result<()> {
16        self.gdn_carry_flush_pending()?;
17        self.release_pools()
18    }
19
20    /// 2026-09-25: Poll this model's InnerQ calibration driver, if it has one.
21    /// A `maybe_finalize` error is logged as a warning and not returned.
22    #[cfg(feature = "cuda")]
23    fn poll_innerq(&self) {
24        if let Some(driver) = self.innerq.as_ref()
25            && let Err(e) = driver.maybe_finalize(128)
26        {
27            tracing::warn!(target: "metrale_model_engine::model::trait_impl", "InnerQ maybe_finalize failed: {e:#}");
28        }
29    }
30
31    fn high_speed_swap_dims(&self) -> Option<metrale_storage::ModelDims> {
32        self.high_speed_swap_dims_dispatch()
33    }
34
35    fn bind_gpu_to_thread(&self) -> Result<()> {
36        self.bind_gpu_to_thread_dispatch()
37    }
38
39    fn alloc_sequence(&self) -> Result<SequenceState> {
40        self.gdn_carry_flush_pending()?;
41        self.alloc_sequence_dispatch(usize::MAX)
42    }
43
44    fn alloc_sequence_for(&self, budget_tokens: usize) -> Result<SequenceState> {
45        self.gdn_carry_flush_pending()?;
46        self.alloc_sequence_dispatch(budget_tokens)
47    }
48
49    fn cache_sequence(&self, seq: &SequenceState) {
50        if let Err(e) = self.gdn_carry_flush_pending() {
51            tracing::error!("gdn carry flush before {}: {e:#}", "cache_sequence");
52        }
53        self.cache_sequence_dispatch(seq)
54    }
55
56    fn free_sequence(&self, seq: &mut SequenceState) -> Result<()> {
57        self.gdn_carry_flush_pending()?;
58        self.free_sequence_dispatch(seq)
59    }
60
61    fn compact_sequence(&self, seq: &mut SequenceState, new_slot: usize) -> Result<()> {
62        self.gdn_carry_flush_pending()?;
63        self.compact_sequence_dispatch(seq, new_slot)
64    }
65
66    fn detach_slot_for_reuse(&self, seq: &mut SequenceState) {
67        if let Err(e) = self.gdn_carry_flush_pending() {
68            tracing::error!("gdn carry flush before {}: {e:#}", "detach_slot_for_reuse");
69        }
70        self.detach_slot_for_reuse_dispatch(seq)
71    }
72
73    fn save_sequence_state(
74        &self,
75        seq: &SequenceState,
76        writer: &mut dyn std::io::Write,
77    ) -> Result<()> {
78        self.gdn_carry_flush_pending()?;
79        self.save_sequence_state_dispatch(seq, writer)
80    }
81
82    fn restore_sequence_state(
83        &self,
84        seq: &mut SequenceState,
85        num_blocks: usize,
86        reader: &mut dyn std::io::Read,
87    ) -> Result<()> {
88        self.gdn_carry_flush_pending()?;
89        self.restore_sequence_state_dispatch(seq, num_blocks, reader)
90    }
91
92    fn num_free_blocks(&self) -> usize {
93        self.num_free_blocks_dispatch()
94    }
95
96    fn num_total_blocks(&self) -> usize {
97        self.num_total_blocks_dispatch()
98    }
99
100    fn kv_block_bytes(&self) -> usize {
101        self.kv_block_bytes_dispatch()
102    }
103
104    fn reclaim_prefix_blocks(&self, num_blocks: usize) -> usize {
105        self.reclaim_prefix_blocks_dispatch(num_blocks)
106    }
107
108    fn reclaim_prefix_blocks_for(
109        &self,
110        prompt: &[u32],
111        adapter_slot: i32,
112        blocks_needed: usize,
113    ) -> crate::traits::PrefixReclaim {
114        self.reclaim_prefix_blocks_for_dispatch(prompt, adapter_slot, blocks_needed)
115    }
116}
metrale_model_engine/model/trait_impl/impl_lifecycle.rs

metrale_model_engine/model/trait_impl/
impl_lifecycle.rs
