fix(qwen35): chunk dense-TP prefill at the rank's prefill batch scaled by the TP degree

crates/hipfire-arch-qwen35/map.md

@@ -40,11 +40,11 @@ _Generated by `scripts/check-crate-maps.py` from the tree — do not edit inside
 | [`src/qwen35/batch.rs`](src/qwen35/batch.rs) | 1,665 | 16 | 0 |
 | [`src/qwen35/config.rs`](src/qwen35/config.rs) | 1,630 | 40 | 21 |
 | [`src/qwen35/ep_batch.rs`](src/qwen35/ep_batch.rs) | 4,800 | 20 | 7 |
-| [`src/qwen35/forward.rs`](src/qwen35/forward.rs) | 6,255 | 31 | 12 |
+| [`src/qwen35/forward.rs`](src/qwen35/forward.rs) | 6,258 | 31 | 12 |
 | [`src/qwen35/load.rs`](src/qwen35/load.rs) | 4,906 | 10 | 0 |
-| [`src/qwen35/prefill.rs`](src/qwen35/prefill.rs) | 9,312 | 11 | 48 |
+| [`src/qwen35/prefill.rs`](src/qwen35/prefill.rs) | 9,332 | 12 | 48 |
 | [`src/qwen35/weights.rs`](src/qwen35/weights.rs) | 1,971 | 43 | 10 |
-| [`src/qwen35.rs`](src/qwen35.rs) | 63 | 7 | 0 |
+| [`src/qwen35.rs`](src/qwen35.rs) | 64 | 7 | 0 |
 | [`src/scheduler.rs`](src/scheduler.rs) | 142 | 3 | 4 |
 | [`src/serve_engine.rs`](src/serve_engine.rs) | 1,273 | 8 | 2 |
 | [`src/slot_batch.rs`](src/slot_batch.rs) | 123 | 4 | 6 |
@@ -74,7 +74,7 @@ _Generated by `scripts/check-crate-maps.py` from the tree — do not edit inside
 - [`src/qwen35/ep_batch.rs`](src/qwen35/ep_batch.rs): `validate_ep_batch_compatibility`, `Qwen35DecodeBatchEpState`, `max_batch`, `lane_capacity`, `epoch`, `poison_mask`, `lane_state`, `new`, `reset_all`, `reset_lane`, `prefill_lane`, `forward_tick`, +8 more
 - [`src/qwen35/forward.rs`](src/qwen35/forward.rs): `dump_expert_stats`, `forward`, `Qwen35Scratch`, `new`, `new_with_kv_max`, `free_gpu`, `Qwen35ScratchSet`, `new_with_kv_max_multi`, `free_gpu_multi`, `forward_scratch`, `prepare_scratch_inputs`, `forward_scratch_with_hidden`, +19 more
 - [`src/qwen35/load.rs`](src/qwen35/load.rs): `hipfire_runtime`, `load_weights`, `HfqSource`, `new`, `ParoSource`, `preflight_weights_dense_tp`, `load_weights_dense_tp_rank`, `set_ep_expert_shard`, `EpShardGuard`, `load_weights_ep_rank`
-- [`src/qwen35/prefill.rs`](src/qwen35/prefill.rs): `PREFILL_MAX_BATCH`, `prefill_max_batch`, `upload_prefill_batch_inputs`, `forward_prefill_batch_single_chunk_captured`, `forward_prefill_batch_single_chunk_captured_opts`, `forward_prefill_batch`, `forward_prefill_batch_capped`, `forward_prefill_batch_with_pbs`, `forward_prefill_batch_with_pbs_opts`, `qwen35_layer_batch_admissible`, `prefill_batch_pbs_eligible`
+- [`src/qwen35/prefill.rs`](src/qwen35/prefill.rs): `PREFILL_MAX_BATCH`, `prefill_max_batch`, `prefill_max_batch_tp`, `upload_prefill_batch_inputs`, `forward_prefill_batch_single_chunk_captured`, `forward_prefill_batch_single_chunk_captured_opts`, `forward_prefill_batch`, `forward_prefill_batch_capped`, `forward_prefill_batch_with_pbs`, `forward_prefill_batch_with_pbs_opts`, `qwen35_layer_batch_admissible`, `prefill_batch_pbs_eligible`
 - [`src/qwen35/weights.rs`](src/qwen35/weights.rs): `DeltaNetLayerWeights`, `FullAttnLayerWeights`, `ExpertWeights`, `mixed_expert_tag`, `SharedExpertWeights`, `MoeFfnWeights`, `MoeParoSidecars`, `DeltaNetMoeLayerWeights`, `FullAttnMoeLayerWeights`, `LayerWeights`, `Qwen35HfqSourceIdentity`, `capture`, +31 more
 - [`src/qwen35.rs`](src/qwen35.rs): `batch`, `config`, `ep_batch`, `forward`, `load`, `prefill`, `weights`
 - [`src/scheduler.rs`](src/scheduler.rs): `Scheduler`, `PendingWork`, `next_batch`
@@ -97,6 +97,6 @@ _Generated by `scripts/check-crate-maps.py` from the tree — do not edit inside
 
 ### Totals
 
-- 29 modules · 57,281 lines · 436 public items · 189 tests · 4 examples
+- 29 modules · 57,305 lines · 437 public items · 189 tests · 4 examples
 
 <!-- crate-map:generated:end -->

crates/hipfire-arch-qwen35/src/qwen35.rs

@@ -46,7 +46,8 @@ pub use prefill::{
     forward_prefill_batch, forward_prefill_batch_capped,
     forward_prefill_batch_single_chunk_captured, forward_prefill_batch_single_chunk_captured_opts,
     forward_prefill_batch_with_pbs, forward_prefill_batch_with_pbs_opts,
-    prefill_batch_pbs_eligible, prefill_max_batch, qwen35_layer_batch_admissible,
+    prefill_batch_pbs_eligible, prefill_max_batch, prefill_max_batch_tp,
+    qwen35_layer_batch_admissible,
     upload_prefill_batch_inputs, PREFILL_MAX_BATCH,
 };
 pub(crate) use prefill::{

crates/hipfire-arch-qwen35/src/qwen35/forward.rs

@@ -4308,7 +4308,10 @@ pub fn forward_prefill_dense_tp(
             _ => return Err(HipError::new(0, "dense TP received a MoE/mismatched layer")),
         }
     }
-    let cap = crate::qwen35::prefill::prefill_max_batch(&gpus.devices[0]);
+    // Size the scratch to the call, not to the arch ceiling: an 8-token verify
+    // would otherwise allocate and free a 512-row PBS per rank on every cycle.
+    let cap = crate::qwen35::prefill::prefill_max_batch_tp(&gpus.devices[0], tp)
+        .min(tokens.len().max(1));
     if cap == 0 {
         return Err(HipError::new(0, "prefill_max_batch is zero"));
     }

crates/hipfire-arch-qwen35/src/qwen35/prefill.rs

@@ -420,6 +420,26 @@ pub fn prefill_max_batch(gpu: &Gpu) -> usize {
     explicit_prefill_max_batch().unwrap_or_else(|| prefill_max_batch_for_arch(gpu.arch.as_str()))
 }
 
+/// Ceiling on the TP compensation below; beyond it the prefill kernels stop
+/// gaining and the per-chunk scratch keeps growing.
+const PREFILL_TP_BATCH_CAP: usize = 2048;
+
+/// Prefill chunk for one rank of a `tp`-way split.
+///
+/// The prefill attention grid is `local_heads x chunk / M_TILE`, and TP divides
+/// the heads, so a rank running the arch default launches `tp` times fewer
+/// workgroups than a single card and starves an already latency-bound kernel.
+/// Scaling the chunk by `tp` restores the single-card workgroup count.
+/// Measured on gfx1100, Qwen3.8-27B, 33k prompt, tp=2: 649 -> 773 tok/s with
+/// byte-identical output. An explicit `HIPFIRE_PREFILL_MAX_BATCH` wins.
+pub fn prefill_max_batch_tp(gpu: &Gpu, tp: usize) -> usize {
+    if let Some(explicit) = explicit_prefill_max_batch() {
+        return explicit;
+    }
+    let base = prefill_max_batch_for_arch(gpu.arch.as_str());
+    base.saturating_mul(tp.max(1)).min(PREFILL_TP_BATCH_CAP)
+}
+
 /// Effective per-chunk capacity for one prefill call.
 ///
 /// Never form a chunk larger than the configured/capped max, the PBS

crates/hipfire-generate/map.md

@@ -27,7 +27,7 @@ _Generated by `scripts/check-crate-maps.py` from the tree — do not edit inside
 | [`src/common.rs`](src/common.rs) | 1,569 | 48 | 1 |
 | [`src/dense.rs`](src/dense.rs) | 8,455 | 91 | 4 |
 | [`src/lib.rs`](src/lib.rs) | 58 | 7 | 0 |
-| [`src/qwen.rs`](src/qwen.rs) | 6,312 | 60 | 1 |
+| [`src/qwen.rs`](src/qwen.rs) | 6,333 | 60 | 1 |
 | [`src/redline.rs`](src/redline.rs) | 4,361 | 49 | 1 |
 | [`src/vision.rs`](src/vision.rs) | 2,942 | 9 | 8 |
 
@@ -55,6 +55,6 @@ _Generated by `scripts/check-crate-maps.py` from the tree — do not edit inside
 
 ### Totals
 
-- 8 modules · 31,785 lines · 321 public items · 224 tests · 0 examples
+- 8 modules · 31,806 lines · 321 public items · 224 tests · 0 examples
 
 <!-- crate-map:generated:end -->

crates/hipfire-generate/src/qwen.rs

@@ -361,7 +361,28 @@ pub fn ep_serve_qwen35_dense_tp(
     );
 
     let t_prefill = Instant::now();
-    for (chunk_index, chunk) in prompt_ids.chunks(32).enumerate() {
+    // Chunk at the rank's prefill batch, not a fixed 32: M=32 misses the MMQ
+    // batch floor (128) and fires the per-layer collectives 16x more often
+    // than the chunk the TP prefill re-chunks to internally (8k prompt on
+    // 2x gfx1100: 72.1 -> 54.7 s wall).
+    let tp_prefill_chunk = match m.ep.as_ref() {
+        Some(EpState { gpus, .. }) => {
+            qwen35::prefill_max_batch_tp(&gpus.devices[0], gpus.devices.len()).max(1)
+        }
+        None => {
+            emit_active_attempt_error(
+                stdout,
+                Some(id),
+                "dense TP serve without EP state",
+                "validation",
+                false,
+                false,
+            );
+            let _ = stdout.flush();
+            return;
+        }
+    };
+    for (chunk_index, chunk) in prompt_ids.chunks(tp_prefill_chunk).enumerate() {
         if check_abort(id) {
             ep_emit_abort(stdout, id, m, 0);
             return;
@@ -395,7 +416,7 @@ pub fn ep_serve_qwen35_dense_tp(
                 weights,
                 configs,
                 chunk,
-                chunk_index * 32,
+                chunk_index * tp_prefill_chunk,
                 kv_caches,
                 dn_states,
                 scratches,