fix(gemma4): enable lowered 26B-A4B generation with row-safe HFQ4-G128 and rollover-safe bounded Q8 KV

crates/hipfire-arch-gemma4/map.md

@@ -31,7 +31,7 @@ _Generated by `scripts/check-crate-maps.py` from the tree — do not edit inside
 | [`src/gemma4.rs`](src/gemma4.rs) | 1,088 | 13 | 0 |
 | [`src/gemma4_vision.rs`](src/gemma4_vision.rs) | 16 | 3 | 0 |
 | [`src/lib.rs`](src/lib.rs) | 51 | 8 | 0 |
-| [`src/lowered.rs`](src/lowered.rs) | 7,427 | 41 | 8 |
+| [`src/lowered.rs`](src/lowered.rs) | 7,457 | 41 | 9 |
 | [`src/speculative.rs`](src/speculative.rs) | 252 | 6 | 0 |
 
 ### Public API surface
@@ -60,6 +60,6 @@ _Generated by `scripts/check-crate-maps.py` from the tree — do not edit inside
 
 ### Totals
 
-- 10 modules · 13,923 lines · 120 public items · 31 tests · 8 examples
+- 10 modules · 13,953 lines · 120 public items · 32 tests · 8 examples
 
 <!-- crate-map:generated:end -->

crates/hipfire-arch-gemma4/src/lowered.rs

@@ -2179,12 +2179,21 @@ pub const GEMMA4_FLASH_TILE: usize = 128;
 /// scale linearly with this.
 pub const GEMMA4_MAX_PREFILL_BATCH: usize = 128;
 
+#[inline]
+fn flash_partials_len_for_tile(
+    max_seq: usize,
+    n_heads: usize,
+    head_dim: usize,
+    tile_size: usize,
+) -> usize {
+    n_heads * max_seq.div_ceil(tile_size) * (2 + head_dim)
+}
+
 /// Pure geometry: single-query flash partial length for `max_seq`.
 /// `n_heads * ceil(max_seq / TILE) * (2 + head_dim)` floats.
 #[inline]
 pub fn gemma4_flash_partials_len(max_seq: usize, n_heads: usize, full_head_dim: usize) -> usize {
-    let tiles = max_seq.div_ceil(GEMMA4_FLASH_TILE);
-    n_heads * tiles * (2 + full_head_dim)
+    flash_partials_len_for_tile(max_seq, n_heads, full_head_dim, GEMMA4_FLASH_TILE)
 }
 
 /// Pure geometry: batched flash partial length for `max_seq`.
@@ -2443,13 +2452,24 @@ impl Gemma4Scratch {
         let i_sample_buf = alloc!(&[2], DType::F32);
         let i_repeat_buf = alloc!(&[1024], DType::F32);
 
-        // Flash partials sizing. Per-head × max_tiles × (2 + head_dim) floats.
-        // Sized for FULL attn (head_dim=512 stride 514, vs sliding 256 stride 258);
-        // sliding-layer dispatches use part of the buffer, full-layer dispatches
-        // use all of it. `max_seq` is the single authority shared with both KV
-        // caches — no independent `HIPFIRE_KV_SEQ` env var.
-        let flash_partials_sz =
+        // gfx1100 sliding Q8 uses a smaller tile than full Asym3 attention, so
+        // its larger tile count can outweigh the smaller per-tile stride.
+        let sliding_tile = rdna_compute::attention::q8_flash_tile_size(
+            &gpu.arch,
+            config.n_heads,
+            config.sliding_n_kv_heads,
+            config.sliding_head_dim,
+            max_seq,
+        );
+        let sliding_flash_partials_sz = flash_partials_len_for_tile(
+            max_seq,
+            config.n_heads,
+            config.sliding_head_dim,
+            sliding_tile,
+        );
+        let full_flash_partials_sz =
             gemma4_flash_partials_len(max_seq, config.n_heads, config.full_head_dim);
+        let flash_partials_sz = sliding_flash_partials_sz.max(full_flash_partials_sz);
         let i_flash_partials = alloc!(&[flash_partials_sz], DType::F32);
 
         // (Note 2026-05-19): removed the precomputed sliding/full cos+sin
@@ -2671,6 +2691,16 @@ impl Gemma4Scratch {
 mod scratch_geometry_tests {
     use super::*;
 
+    #[test]
+    fn sliding_tile32_partials_can_exceed_full_tile128_partials() {
+        let max_seq = 1025;
+        let sliding = flash_partials_len_for_tile(max_seq, 16, 256, 32);
+        let full = flash_partials_len_for_tile(max_seq, 16, 512, 128);
+        assert!(sliding > full);
+        assert_eq!(sliding, 16 * 33 * 258);
+        assert_eq!(full, 16 * 9 * 514);
+    }
+
     fn dummy_cfg_31b() -> Gemma4Config {
         // Minimal config mirroring 31B/26B shapes: n_heads=32, full_head_dim=512
         Gemma4Config {

crates/hipfire-dispatch/map.md

@@ -24,7 +24,7 @@ _Generated by `scripts/check-crate-maps.py` from the tree — do not edit inside
 |---|---:|---:|---:|
 | [`src/context.rs`](src/context.rs) | 67 | 5 | 0 |
 | [`src/coverage_tests.rs`](src/coverage_tests.rs) | 1,841 | 0 | 21 |
-| [`src/families/attention.rs`](src/families/attention.rs) | 2,463 | 9 | 10 |
+| [`src/families/attention.rs`](src/families/attention.rs) | 2,471 | 9 | 10 |
 | [`src/families/fused_qkv.rs`](src/families/fused_qkv.rs) | 1,579 | 8 | 1 |
 | [`src/families/gemm.rs`](src/families/gemm.rs) | 647 | 7 | 2 |
 | [`src/families/gemv.rs`](src/families/gemv.rs) | 608 | 20 | 0 |
@@ -105,6 +105,6 @@ _Generated by `scripts/check-crate-maps.py` from the tree — do not edit inside
 
 ### Totals
 
-- 33 modules · 22,681 lines · 223 public items · 242 tests · 0 examples
+- 33 modules · 22,689 lines · 223 public items · 242 tests · 0 examples
 
 <!-- crate-map:generated:end -->

crates/hipfire-dispatch/src/families/attention.rs

@@ -1187,6 +1187,14 @@ fn dispatch_attend(
                 let st = io.givens_sin.unwrap();
                 let fp = io.flash_partials.unwrap();
                 if io.head_dim == 512 {
+                    if io.output_gate.is_some() {
+                        return Err(DispatchError::UnsupportedVariant {
+                            family: "attention/attend",
+                            variant: "D512 Asym3 output gate is unsupported",
+                            arch: "",
+                            quant: "asym3",
+                        });
+                    }
                     return hip!(gpu.attention_flash_asym3_hd512(
                         io.q,
                         io.k_cache,