prepare, levers: a checkpoint's BF16 matrices stay BF16 by default; their Q8_0 rewrite behind BLOOMERY_PREPARE_Q8

crates/hf/src/modelopt.rs

@@ -1093,6 +1093,11 @@ mod tests {
         );
         let kept = normalize(get, &[("b.weight", &b)], Use::Matrix(Plain::Bf16)).unwrap();
         assert_eq!((kept.type_id, &kept.bytes), (30, &bf16));
+        let f32 = placed(Format::F32, &[]);
+        let kept = normalize(get, &[("f.weight", &f32)], Use::Matrix(Plain::F32)).unwrap();
+        assert_eq!((kept.type_id, &kept.bytes), (0, &f32s));
+        let q = normalize(get, &[("f.weight", &f32)], m).unwrap();
+        assert_eq!((q.type_id, &q.bytes), (8, &want));
         let rows = normalize(
             get,
             &[("t.weight", &placed(Format::Fp8Tensor, &["t.scale"]))],

crates/levers/src/lib.rs

@@ -39,7 +39,7 @@ pub use registry::{
     PREFILL_GROUP_DEFAULT, PREFILL_GROUP_MAX, QWEN3_KV, QWEN38_EXPERTS, R8, RESIDENCY, ROUTE_TRACE,
     SCREEN, SPIN, STEP_STATS, THREADS, TIER_LINK, TIER_NODES, TP_NODES, WARM_PROFILE,
 };
-pub use registry::{SCHEMAJUMP, SPLIT_WALK, XSTREAM};
+pub use registry::{PREPARE_Q8, SCHEMAJUMP, SPLIT_WALK, XSTREAM};
 
 #[cfg(test)]
 mod tests;
@@ -1183,6 +1183,13 @@ impl Levers {
         self.flag(SCHEMAJUMP)
     }
 
+    /// `BLOOMERY_PREPARE_Q8`: `prepare` writes the BF16 or F32 matrices its map
+    /// reads as Q8_0 in Q8_0, not as stored.
+    #[must_use]
+    pub fn prepare_q8(&self) -> bool {
+        self.flag(PREPARE_Q8)
+    }
+
     /// The host tier's load settings: [`HOST_POPULATE`], [`HOST_LOCK`],
     /// [`CARD_DONTNEED`], [`R8`], and whether [`ROUTE_TRACE`] is set.
     #[must_use]
@@ -2025,7 +2032,7 @@ pub enum Draft38Off {
     /// Unset, and the last window needs `need` positions of `ctx`.
     Ctx { need: usize, ctx: usize },
     /// Unset on the Qwen3.8 seat, and the target's `name`, a matrix the
-    /// draft borrows, is `ty` (or `absent`), not the Q8_0 the draft reads.
+    /// draft borrows, is `ty` (or `absent`), not a form the draft reads.
     Borrowed { name: String, ty: String },
     /// Unset, and the draft yielded to the context
     /// ([`RungCost::of`]'s answer, the seats' plan-side rule).
@@ -2072,7 +2079,7 @@ impl fmt::Display for Draft38Off {
             ),
             Draft38Off::Borrowed { name, ty } => write!(
                 f,
-                "unset: the target's {name} is {ty}; the MTP draft reads it as Q8_0"
+                "unset: the target's {name} is {ty}, not a form the MTP draft reads"
             ),
             Draft38Off::Yield(y) => y.fmt(f),
             Draft38Off::Paged { arena } => write!(

crates/levers/src/registry.rs

@@ -51,6 +51,7 @@ pub const TIER_NODES: &str = "BLOOMERY_TIER_NODES";
 pub const TP_NODES: &str = "BLOOMERY_TP_NODES";
 pub const SCREEN: &str = "BLOOMERY_SCREEN";
 pub const SCHEMAJUMP: &str = "BLOOMERY_SCHEMAJUMP";
+pub const PREPARE_Q8: &str = "BLOOMERY_PREPARE_Q8";
 
 /// The largest `BLOOMERY_PREFILL_GROUP`: the batches a V4.1 prompt group
 /// holds at most, which the body's buffers are sized for.
@@ -1073,6 +1074,30 @@ pub(crate) static REGISTRY: &[LeverSpec] = &[
             phase: Phase::Load,
         },
     },
+    LeverSpec {
+        name: PREPARE_Q8,
+        class: Class::M,
+        kind: Kind::Flag,
+        default: Unset::Is("0"),
+        doc: "`prepare`: `1` writes each BF16 or F32 matrix its checkpoint map reads as Q8_0 (the \
+              attention and linear-attention projections, the shared expert, the PLE key and \
+              value, the hyper-connections' down and up, the head and the token embedding) as \
+              ggml's reference Q8_0; `0` keeps each single BF16 matrix as stored and refuses \
+              F32 matrices read as Q8_0. BF16 and F32 expert stacks are refused, and the \
+              quantized tensors (NVFP4, MXFP8, FP8) keep their stored blocks, under both \
+              values. The rewrite is lossy: on \
+              `nvidia/Qwen3.8-Flash-Next-NVFP4` at fc694b54, the checkpoint's own f32 forward \
+              with its BF16 projections rounded to Q8_0 is at a KL divergence of 4.288e-02 from \
+              the forward with them as stored, at the last of the five positions of the \
+              batch set, and 4.181e-02 at position 3,000 of the D3K set (refset \
+              `hf-qwen4exp-nvfp4`, the `_bf16` sets against the others), so it is asked for, \
+              never a default. `--keep-bf16`, which asks for the stored form, is refused \
+              beside it.",
+        site: Site::Parsed {
+            left: &[],
+            phase: Phase::Load,
+        },
+    },
     LeverSpec {
         name: "BLOOMERY_POISON",
         class: Class::D,

crates/levers/src/tests.rs

@@ -404,6 +404,7 @@ fn accessors_read_their_rows() {
     assert_eq!(unset.host_lanes(), HOST_LANES_DEFAULT);
     assert_eq!(unset.nvtier_mem(), NvTierMem::Auto);
     assert_eq!(unset.gen_slots(), 1);
+    assert!(!unset.prepare_q8());
     assert_eq!(
         unset.host(),
         HostCfg {
@@ -448,6 +449,7 @@ fn accessors_read_their_rows() {
             (NVTIER_MEM, "pinned"),
             (PREFILL_GROUP, "8"),
             (GEN_SLOTS, "2"),
+            (PREPARE_Q8, "1"),
         ]),
         Scope::Every,
     )
@@ -498,6 +500,7 @@ fn accessors_read_their_rows() {
         );
     }
     assert_eq!(set.gen_slots(), 2);
+    assert!(set.prepare_q8());
     assert_eq!(
         set.host(),
         HostCfg {
@@ -553,7 +556,8 @@ fn accessors_read_their_rows() {
             GEN_SLOTS,
             FORCE_IDS,
             SCREEN,
-            SCHEMAJUMP
+            SCHEMAJUMP,
+            PREPARE_Q8
         ]
     );
 }
@@ -1303,10 +1307,10 @@ fn draft38_unset_follows_the_place_and_the_file() {
     assert_eq!(
         Draft38Off::Borrowed {
             name: "output.weight".to_string(),
-            ty: "q6_K".to_string()
+            ty: "f16".to_string()
         }
         .to_string(),
-        "unset: the target's output.weight is q6_K; the MTP draft reads it as Q8_0"
+        "unset: the target's output.weight is f16, not a form the MTP draft reads"
     );
 }
 
@@ -1449,7 +1453,7 @@ fn a_paged_plan_turns_only_an_unset_draft_off() {
         Draft38Off::Ctx { need: 5, ctx: 4 },
         Draft38Off::Borrowed {
             name: "output.weight".to_string(),
-            ty: "q6_K".to_string(),
+            ty: "f16".to_string(),
         },
     ] {
         assert_eq!(

crates/model-hf/Cargo.toml

@@ -15,6 +15,8 @@ workspace = true
 [dependencies]
 bloomery-gguf = { path = "../gguf" }
 bloomery-hf = { path = "../hf" }
+# The `prepare` bin reads its one lever, `BLOOMERY_PREPARE_Q8`.
+bloomery-levers = { path = "../levers" }
 bloomery-model = { path = "../model" }
 bloomery-safetensors = { path = "../safetensors" }
 # The `prepare` bin builds serving's tokenizer from the GGUF header it wrote and checks its split.