perf: reduce decode overhead in scheduler and executor

- Precompute page_table and decode_mask on KVCache once per step in PagePool.bind_tasks, instead of per-layer in CudaBackend/TorchNativeBackend
- Skip frequency penalty history tensor construction when all penalties are 0 in Executor.execute_decode
- Omit FrequencyPenaltyStrategy from sampling pipeline when penalty is 0
- Deduplicate get_active_tasks calls in scheduler loop (3 to 1), remove redundant sorted() on decode tasks
- Benchmark (L20, bf16, CUDA backend): B=1 9.48->9.40ms (+1%), B=4 10.73->9.89ms (+8.6%), B=8 10.77->10.13ms (+6.4%)

astrai/inference/core/cache.py

@@ -203,6 +203,10 @@ class KVCache:
         seq_lens: [batch_size] — per-request total sequence lengths
         out_cache_loc: [batch, new_seq_len] or [batch, 1] — write indices
         max_len: max(seq_lens) as Python int — avoids GPU sync in decode
+        page_table: [batch, max_len] — precomputed gather indices for decode;
+            None for prefill or when not yet computed.
+        decode_mask: [batch, max_len] bool — precomputed position validity
+            mask for decode; None for prefill or single-batch decode.
     """
 
     k_buffer: Tensor
@@ -212,6 +216,8 @@ class KVCache:
     seq_lens: Tensor
     out_cache_loc: Tensor
     max_len: int = 0
+    page_table: Optional[Tensor] = None
+    decode_mask: Optional[Tensor] = None
 
 
 class PagePool:
@@ -428,11 +434,21 @@ class PagePool:
             out_cache_loc = self._req_pool.req_to_token[
                 req_pool_indices, start_pos:seq_len
             ]
+            page_table = None
+            decode_mask = None
         else:
             write_pos = seq_lens_t - 1
             out_cache_loc = self._req_pool.req_to_token[
                 req_pool_indices, write_pos
             ].unsqueeze(-1)
+            ml = max(seq_lens)
+            page_table = self._req_pool.req_to_token[req_pool_indices, :ml]
+            if len(task_ids) > 1:
+                decode_mask = (
+                    torch.arange(ml, device=device)[None, :] < seq_lens_t[:, None]
+                )
+            else:
+                decode_mask = None
 
         return KVCache(
             k_buffer=self._storage.k_buffer,
@@ -442,6 +458,8 @@ class PagePool:
             seq_lens=seq_lens_t,
             out_cache_loc=out_cache_loc,
             max_len=max(seq_lens),
+            page_table=page_table,
+            decode_mask=decode_mask,
         )
 
     # ---- internals ----

astrai/inference/core/executor.py

@@ -105,26 +105,33 @@ class Executor:
             [t.frequency_penalty for t in tasks], device=self.device
         )
 
-        history_lists = []
-        history_lens = []
-        for t in tasks:
-            window = t.rep_window
-            prompt_part = t.prompt_ids[-window:]
-            ids = prompt_part + t.output_ids
-            history_lists.append(ids)
-            history_lens.append(len(ids))
+        has_freq = bool((freq_penalties != 0).any())
+        if has_freq:
+            history_lists = []
+            history_lens = []
+            for t in tasks:
+                window = t.rep_window
+                prompt_part = t.prompt_ids[-window:]
+                ids = prompt_part + t.output_ids
+                history_lists.append(ids)
+                history_lens.append(len(ids))
 
-        max_len = max(history_lens) if history_lens else 0
-        padded_ids = torch.zeros(
-            len(tasks), max_len, dtype=torch.long, device=self.device
-        )
-        padded_mask = torch.zeros(
-            len(tasks), max_len, dtype=torch.bool, device=self.device
-        )
-        for i, h in enumerate(history_lists):
-            L = history_lens[i]
-            padded_ids[i, :L] = torch.as_tensor(h, dtype=torch.long, device=self.device)
-            padded_mask[i, :L] = True
+            max_len = max(history_lens) if history_lens else 0
+            padded_ids = torch.zeros(
+                len(tasks), max_len, dtype=torch.long, device=self.device
+            )
+            padded_mask = torch.zeros(
+                len(tasks), max_len, dtype=torch.bool, device=self.device
+            )
+            for i, h in enumerate(history_lists):
+                L = history_lens[i]
+                padded_ids[i, :L] = torch.as_tensor(
+                    h, dtype=torch.long, device=self.device
+                )
+                padded_mask[i, :L] = True
+        else:
+            padded_ids = None
+            padded_mask = None
 
         with torch.inference_mode():
             outputs = self.model(

astrai/inference/core/scheduler.py

@@ -109,9 +109,11 @@ class InferenceScheduler:
                     self._task_mgr.wait_for_tasks(timeout=1.0)
                     continue
 
+                active = self._task_mgr.get_active_tasks()
+
                 to_prefill = [
                     t
-                    for t in self._task_mgr.get_active_tasks()
+                    for t in active
                     if t.output_tokens == 0
                     and cache.task_cached(t.task_id) < len(t.prompt_ids)
                 ]
@@ -137,10 +139,10 @@ class InferenceScheduler:
                                 t.task_id, t.prompt_ids, start_logical_page
                             )
 
-                decode_tasks = self._task_mgr.get_active_tasks()
+                decode_tasks = active
 
                 valid: List[Task] = []
-                for t in sorted(decode_tasks, key=lambda t: t.task_id):
+                for t in decode_tasks:
                     if cache.task_extend(t.task_id, t.next_pos):
                         valid.append(t)
                     else: