Move string prefill overloads to .cpp, fix stale prefill_next_token_

Move the two string convenience prefill methods from inline in
text_llm_runner.h to text_llm_runner.cpp. This removes the header's
dependency on multimodal_input.h — only a forward declaration of
MultimodalInput (from irunner.h) is needed in the header.

Also fix a bug where MultimodalRunner::generate(vector) left
prefill_next_token_ set after consuming the token. A subsequent
generate("") would incorrectly reuse the stale token. Clear it
after extracting the value.

This PR was authored with the assistance of Claude.

Author: kirklandsign

extension/llm/runner/multimodal_runner.cpp

@@ -304,6 +304,7 @@ Error MultimodalRunner::generate(
   auto prefill_result = prefill(inputs, config.num_bos, config.num_eos);
   ET_CHECK_OK_OR_RETURN_ERROR(prefill_result.error());
   uint64_t cur_token = prefill_result.get();
+  prefill_next_token_.reset();
 
   return decode_from_token(cur_token, config, wrapped_callback, stats_callback);
 }

extension/llm/runner/text_llm_runner.cpp

@@ -11,6 +11,7 @@
 // The module takes in a string as input and emits a string as output.
 
 #include <executorch/extension/llm/runner/io_manager/io_manager.h>
+#include <executorch/extension/llm/runner/multimodal_input.h>
 #include <executorch/extension/llm/runner/text_llm_runner.h>
 #include <executorch/extension/llm/runner/util.h>
 #include <executorch/runtime/platform/runtime.h>
@@ -289,6 +290,21 @@ Result<uint64_t> TextLLMRunner::prefill(
   return prefill_next_token_.value();
 }
 
+Result<uint64_t> TextLLMRunner::prefill(
+    const std::string& prompt,
+    int32_t num_bos,
+    int32_t num_eos) {
+  std::vector<MultimodalInput> inputs;
+  inputs.emplace_back(MultimodalInput(prompt));
+  return prefill(inputs, num_bos, num_eos);
+}
+
+Result<uint64_t> TextLLMRunner::prefill(
+    const std::string& prompt,
+    const GenerationConfig& config) {
+  return prefill(prompt, config.num_bos, config.num_eos);
+}
+
 Error TextLLMRunner::warmup(const std::string& prompt, int32_t max_new_tokens) {
   // Create a GenerationConfig for warmup
   GenerationConfig config;

extension/llm/runner/text_llm_runner.h

@@ -19,7 +19,6 @@
 #include <unordered_map>
 
 #include <executorch/extension/llm/runner/irunner.h>
-#include <executorch/extension/llm/runner/multimodal_input.h>
 #include <executorch/extension/llm/runner/stats.h>
 #include <executorch/extension/llm/runner/text_decoder_runner.h>
 #include <executorch/extension/llm/runner/text_prefiller.h>
@@ -120,22 +119,18 @@ class ET_EXPERIMENTAL TextLLMRunner : public IRunner {
   /**
    * Convenience overload: prefill a single text prompt.
    */
-  ::executorch::runtime::Result<uint64_t>
-  prefill(const std::string& prompt, int32_t num_bos = 0, int32_t num_eos = 0) {
-    std::vector<MultimodalInput> inputs;
-    inputs.emplace_back(MultimodalInput(prompt));
-    return prefill(inputs, num_bos, num_eos);
-  }
+  ::executorch::runtime::Result<uint64_t> prefill(
+      const std::string& prompt,
+      int32_t num_bos = 0,
+      int32_t num_eos = 0);
 
   /**
    * Prefill a text prompt using GenerationConfig.
    * Deprecated: prefer prefill(prompt, num_bos, num_eos).
    */
   ::executorch::runtime::Result<uint64_t> prefill(
       const std::string& prompt,
-      const GenerationConfig& config) {
-    return prefill(prompt, config.num_bos, config.num_eos);
-  }
+      const GenerationConfig& config);
 
   /**
    * @brief Warms up the model with a sample prompt