#0: minor fixes

tt_metal/impl/device/device.cpp

@@ -1205,6 +1205,7 @@ bool Device::initialize(const uint8_t num_hw_cqs, size_t l1_small_size, const st
     log_info(tt::LogMetal, "Initializing device {}. Program cache is {}enabled", this->id_, this->program_cache.is_enabled() ? "": "NOT ");
     TT_ASSERT(num_hw_cqs > 0 and num_hw_cqs < 3, "num_hw_cqs can be between 1 and 2");
     this->build_key_ = tt::Cluster::instance().get_harvesting_mask(this->id());
+    this->using_fast_dispatch = false;
     this->num_hw_cqs_ = num_hw_cqs;
     this->initialize_cluster();
     this->initialize_allocator(l1_small_size, l1_bank_remap);

tt_metal/impl/device/device.hpp

@@ -250,7 +250,7 @@ class Device {
     uint8_t num_hw_cqs_;
 
     vector<std::unique_ptr<Program, tt::tt_metal::detail::ProgramDeleter>> command_queue_programs;
-    bool using_fast_dispatch = false;
+    bool using_fast_dispatch;
     program_cache::detail::ProgramCache program_cache;
 
     // Program cache interface. Syncrhonize with worker worker threads before querying or
@@ -264,9 +264,9 @@ class Device {
         log_info(tt::LogMetal, "Disabling and clearing program cache on device {}", this->id_);
         this->synchronize();
         if (this->program_cache.is_enabled()) {
-            program_cache.clear();
             program_cache.disable();
         }
+        program_cache.clear();
     }
     std::size_t num_program_cache_entries() {
         this->synchronize();

tt_metal/impl/dispatch/command_queue.cpp

@@ -944,6 +944,7 @@ void EnqueueProgramCommand::process() {
         this->assemble_device_commands();
     } else {
         static constexpr uint32_t count_offset = (sizeof(CQPrefetchCmd) + offsetof(CQDispatchCmd, wait.count));
+        TT_ASSERT(this->cached_program_command_sequences.find(program.id) != this->cached_program_command_sequences.end(), "Program cache hit, but no stored command sequence");
         this->cached_program_command_sequences[program.id].preamble_command_sequence.update_cmd_sequence(
             count_offset, &this->expected_num_workers_completed, sizeof(uint32_t));
         this->assemble_device_commands();