feat(lora): runtime BF16 LoRA adapter support (Holo-3.1 dense + 35B MoE)

.github/workflows/file-size-cap.yml

@@ -41,6 +41,17 @@ jobs:
             # external contributors. Candidate for a `traits/model/mod.rs`
             # split that pulls verify/decode default impls into siblings.
             "crates/spark-model/src/traits/model.rs"
+            # ── PR: runtime LoRA adapter MVP carry-over ──
+            # 502 LoC — `TransformerLayer` trait: the core per-layer interface
+            # (~20 methods spanning decode / prefill / SSM-phase / MoE-transpose,
+            # each with a long arg list + doc). Atomic interface, same rationale
+            # as `traits/model.rs` above: the SSM-prefill defaults are one-line
+            # `bail!`/`Ok(())` no-ops (no extractable logic) and the arg-heavy
+            # signatures make default-body extraction net-neutral. Tipped over
+            # 500 by the `as_any_mut` downcast hook the LoRA install walk needs.
+            # Clean split (SSM-prefill phases → sibling supertrait) tracked as
+            # follow-up.
+            "crates/spark-model/src/layer/transformer_layer.rs"
             # 513 LoC — Gemma-4 weight loader (variant-dispatch family).
             # Already part of a `gemma4/` split; the leftover holds the
             # single-file dispatch entry-point + per-layer-type orchestration.

.gitignore

@@ -128,3 +128,6 @@ bench/**/*.jsonl
 bench/**/*.csv
 bench/**/*.pid
 bench/fp8_dgx2_drift/harness/oc
+
+# LoRA MVP: large fp32 reference deltas (regenerate via scripts/reference_deltas.py in seconds)
+test_data/lora-holo-tiny*/reference_deltas.safetensors

crates/atlas-core/src/config.rs

@@ -382,6 +382,14 @@ pub struct ModelConfig {
     /// is what the drafter's `fc` projection expects.
     #[serde(default)]
     pub dflash_capture_layers: Vec<usize>,
+
+    /// LoRA adapter rank ceiling (`--max-lora-rank`). `0` = LoRA disabled.
+    /// Set programmatically before model build (never parsed from the HF
+    /// `config.json`); the only consumer is `BufferSizes`, which sizes the
+    /// adapter delta scratch from it. `adapter_*` naming avoids the MLA
+    /// `*lora_rank` collision (`config.rs:182-207`).
+    #[serde(default)]
+    pub adapter_max_rank: usize,
 }
 
 /// Advertised weight-quantization layout, as declared in the HF
@@ -484,10 +492,13 @@ mod parsers;
 mod tests;
 
 pub use dispatch::parse_config;
+pub use parsers::{
+    PEFT_SUPPORTED_TARGET_MODULES, PeftAdapterConfig, parse_mistral_params,
+    parse_peft_adapter_config, parse_quantization_config,
+};
 pub(crate) use parsers::{
     parse_deepseek_v4, parse_gemma4_params, parse_minimax_m2, parse_step3p7, parse_vision_config,
 };
-pub use parsers::{parse_mistral_params, parse_quantization_config};
 
 pub(crate) fn finalize_config(config: &mut ModelConfig, raw: &serde_json::Value) -> Result<()> {
     if config.quantization_config.is_none() {

crates/atlas-core/src/config/factory.rs

@@ -107,6 +107,7 @@ impl ModelConfig {
             mtp_transformer_layers: 0,
             rotary_dim: 0,
             dflash_capture_layers: Vec::new(),
+            adapter_max_rank: 0,
         }
     }
 }