StaticAttetnion runtime support for generate_full_logits=False

Differential Revision: D88790445

Pull Request resolved: #16171

Author: sxu

examples/models/llama/runner/static_attention_io_manager.h

@@ -438,6 +438,8 @@ class StaticAttentionIOManager {
     RopeT* rope_freqs_cos;
     RopeT* rope_freqs_sin;
     StaticAttentionUpdateStyle style = StaticAttentionUpdateStyle::SMART_MASK;
+    bool generate_full_logits = true;
+    std::optional<size_t> last_valid_token_pos_index = 0;
   };
 
   StaticAttentionIOManager(StaticAttentionIOConfig config)
@@ -607,9 +609,16 @@ class StaticAttentionIOManager {
       batch_len = std::min(input_len, tokens.size() - i);
       if (input_pos_ + batch_len > config_.max_context_len) {
         ET_LOG(Error, "Maximum context size reached, stopping prefill.");
-        return input_len - 1;
+        return config_.generate_full_logits ? input_len - 1 : 0;
       }
       std::copy(&tokens[i], &tokens[i + batch_len], input_buffer.begin());
+      if (!config_.generate_full_logits && config_.last_valid_token_pos_index) {
+        last_valid_token_pos_ = batch_len - 1;
+        set_input(
+            method,
+            *config_.last_valid_token_pos_index,
+            &last_valid_token_pos_);
+      }
       prepare(method);
       ET_CHECK(method.execute() == executorch::runtime::Error::Ok);
       update(
@@ -622,10 +631,12 @@ class StaticAttentionIOManager {
         auto* logits = logits_tensor.const_data_ptr<LogitT>();
         logits_callback(executorch::runtime::Span(
             logits,
-            logits + batch_len * logits_tensor.size(logits_tensor.dim() - 1)));
+            logits +
+                (config_.generate_full_logits ? batch_len : 1) *
+                    logits_tensor.size(logits_tensor.dim() - 1)));
       }
     }
-    return batch_len - 1;
+    return config_.generate_full_logits ? batch_len - 1 : 0;
   }
 
   /**
@@ -648,6 +659,11 @@ class StaticAttentionIOManager {
       mask.set_causal_mask();
       set_input(method, config_.cache_len_to_mask_idx[pair.first], mask.get());
     }
+    if (!config_.generate_full_logits && config_.last_valid_token_pos_index) {
+      last_valid_token_pos_ = 0;
+      set_input(
+          method, *config_.last_valid_token_pos_index, &last_valid_token_pos_);
+    }
 
     while (true) {
       input_buffer[0] = prev_tok;
@@ -685,6 +701,7 @@ class StaticAttentionIOManager {
       size_t window_size,
       size_t n_verifications,
       std::unordered_map<TokenT, SuffixCache<TokenT>> suffix_caches) {
+    ET_CHECK(config_.generate_full_logits);
     ET_LOG(
         Info,
         "Decoding with lookahead and verification at position %zu",
@@ -968,6 +985,7 @@ class StaticAttentionIOManager {
   std::unordered_map<size_t, PerCacheLenMasks> attentionMasks_;
   std::vector<RopeT> rope_freqs_cos_override_;
   std::vector<RopeT> rope_freqs_sin_override_;
+  int64_t last_valid_token_pos_;
 };
 
 } // namespace example

examples/models/llama/static_attention.py

@@ -456,7 +456,10 @@ def decode(
         new_tokens = [init_token]
         for _ in range(n):
             y = self._run_once(model, new_tokens[-1:])[0]
-            new_tokens.append(y[:, :1, ...].argmax().item())
+            if self.generate_full_logits:
+                new_tokens.append(y[:, :1, ...].argmax().item())
+            else:
+                new_tokens.append(y.argmax().item())
             if new_tokens[-1] in stop_tokens:
                 break
 
@@ -607,13 +610,20 @@ def _run_once(
             freqs_cos_override = self.freqs_cos[self.pos : self.pos + self.input_len]
         if freqs_sin_override is None:
             freqs_sin_override = self.freqs_sin[self.pos : self.pos + self.input_len]
+        if not self.generate_full_logits:
+            extra_attn_options = {
+                "last_valid_token_pos": torch.tensor([n_tokens - 1], dtype=torch.long)
+            }
+        else:
+            extra_attn_options = {}
         y, attn_updates = model(
             tokens,
             {
                 "masks": self.masks,
                 "freqs_cos_override": freqs_cos_override,
                 "freqs_sin_override": freqs_sin_override,
                 "in_cache_state": (self.k_caches, self.v_caches),
+                **extra_attn_options,
             },
         )
         non_padded_len = non_padded_len or n_tokens

examples/models/llama/tests/test_static_attention.py

@@ -248,7 +248,9 @@ def test(style, attention_type):
                 )
                 ys.append(y_i)
 
-            self.assertTrue(torch.isclose(ys[-1], expected, rtol=1e-3).all())
+            self.assertTrue(
+                torch.isclose(ys[-1].flatten(), expected.flatten(), rtol=1e-3).all()
+            )
 
         for args in itertools.product(
             ["shift_pointer", "smart_mask"], ["static", "static_mha"]