Merge branch 'main' into stable

Author: ufownl

README.md

@@ -114,7 +114,6 @@ Available options:
   weights = "/path/to/4b-it-sfp.sbs",  -- Path of model weights file. (required)
   map = -1,  -- Enable memory-mapping? (-1 means auto, 0 means no, 1 means yes)
   to_bf16 = -1,  -- Convert weights to bf16? (-1 means auto, 0 means no, 1 means yes)
-  seed = 42,  -- Random seed. (default is random setting)
   scheduler = sched_inst,  -- Instance of scheduler, if not provided a default
                            -- scheduler will be attached.
   disabled_words = {...},  -- Words you don't want to generate.

src/batch.cpp

@@ -157,7 +157,6 @@ int batch(lua_State* L) {
     auto cfg = parse_config(sess_ctxs);
     cfg.verbosity = 0;
     auto inst = sess_ctxs.front().sess->inst();
-    cfg.gen = &inst->rnd();
     cfg.batch_stream_token = [&](size_t query_idx, size_t pos, int token, float) {
       auto& ctx = sess_ctxs[query_idx];
       if (ctx.stream_fn == 0) {

src/image_tokens.cpp

@@ -79,7 +79,6 @@ int create(lua_State* L) {
     );
     tks.AllocateAndAttachRowPtrs(inst->matmul_env().row_ptrs);
     gcpp::RuntimeConfig cfg;
-    cfg.gen = &inst->rnd();
     cfg.verbosity = 0;
     inst->model().GenerateImageTokens(cfg, tks.Rows(), img, tks, inst->matmul_env());
     auto ud = lua_newuserdata(L, sizeof(gcpp::ImageTokens));

src/instance.cpp

@@ -32,9 +32,8 @@ int disabled_tokens(lua_State* L) {
 
 namespace cgemma {
 
-instance::instance(int argc, char* argv[], unsigned int seed, scheduler* sched)
+instance::instance(int argc, char* argv[], scheduler* sched)
   : args_(argc, argv)
-  , rnd_(seed)
   , sched_(sched) {
   if (!sched_) {
     default_sched_ = std::make_unique<scheduler>();
@@ -116,19 +115,10 @@ int instance::create(lua_State* L) {
   }
   auto ud = lua_newuserdata(L, sizeof(instance));
   try {
-    unsigned int seed;
-    lua_getfield(L, 1, "seed");
-    if (lua_isnumber(L, -1)) {
-      seed = lua_tointeger(L, -1);
-    } else {
-      std::random_device rd;
-      seed = rd();
-    }
-    lua_pop(L, 1);
     lua_getfield(L, 1, "scheduler");
     auto sched = scheduler::to(L, -1);
     lua_pop(L, 1);
-    auto inst = new(ud) instance(argc, argv, seed, sched);
+    auto inst = new(ud) instance(argc, argv, sched);
     luaL_getmetatable(L, name);
     lua_setmetatable(L, -2);
     lua_getfield(L, 1, "disabled_words");

src/instance.hpp

@@ -5,7 +5,6 @@
 #include <gemma/gemma.h>
 #include <gemma/gemma_args.h>
 #include <unordered_set>
-#include <random>
 #include <memory>
 
 namespace cgemma {
@@ -15,10 +14,9 @@ constexpr const int UNK_ID = 3;
 
 class instance {
 public:
-  instance(int argc, char* argv[], unsigned int seed, scheduler* sched);
+  instance(int argc, char* argv[], scheduler* sched);
 
   const gcpp::LoaderArgs& args() const { return args_; }
-  std::mt19937& rnd() { return rnd_; }
   gcpp::ThreadingContext& threading_ctx() const { return sched_->threading_ctx(); }
   gcpp::MatMulEnv& matmul_env() const { return sched_->matmul_env(); }
   gcpp::Gemma& model() const { return *model_; }
@@ -33,7 +31,6 @@ class instance {
 
 private:
   gcpp::LoaderArgs args_;
-  std::mt19937 rnd_;
   scheduler* sched_;
   std::unique_ptr<scheduler> default_sched_;
   std::unique_ptr<gcpp::Gemma> model_;

src/session.cpp

@@ -15,7 +15,6 @@ void generate(cgemma::session* sess, const gcpp::ImageTokens* image, const std::
   gcpp::RuntimeConfig cfg;
   sess->args().CopyTo(cfg);
   cfg.verbosity = 0;
-  cfg.gen = &sess->inst()->rnd();
   cfg.batch_stream_token = stream_token;
   if (!sess->inst()->disabled_tokens().empty()) {
     cfg.accept_token = [&](int token, float) {
@@ -139,8 +138,6 @@ int reset(lua_State* L) {
 
 enum class kv_cache_field: size_t {
   kv_cache,
-  conv1d_cache,
-  rglru_cache,
   end
 };
 
@@ -158,20 +155,6 @@ class kv_cache_blob {
       ptrs_[static_cast<size_t>(kv_cache_field::kv_cache)] = nullptr;
       sizes_[static_cast<size_t>(kv_cache_field::kv_cache)] = 0;
     }
-#define GRIFFIN_CACHE(FIELD)                                                                                      \
-  do {                                                                                                            \
-    auto& field = sess->kv_cache().FIELD;                                                                         \
-    if (field.Rows() > 0) {                                                                                       \
-      ptrs_[static_cast<size_t>(kv_cache_field::FIELD)] = field.RowBytes(0);                                      \
-      sizes_[static_cast<size_t>(kv_cache_field::FIELD)] = field.Rows() * field.Stride() * field.ElementBytes();  \
-    } else {                                                                                                      \
-      ptrs_[static_cast<size_t>(kv_cache_field::FIELD)] = nullptr;                                                \
-      sizes_[static_cast<size_t>(kv_cache_field::FIELD)] = 0;                                                     \
-    }                                                                                                             \
-  } while (false)
-    GRIFFIN_CACHE(conv1d_cache);
-    GRIFFIN_CACHE(rglru_cache);
-#undef GRIFFIN_CACHE
   }
 
   template <kv_cache_field Field>
@@ -207,8 +190,6 @@ size_t dump_impl(char* buf, const cgemma::session* sess) {
     }                                                                                             \
   } while (false)
     DUMP_CACHE(kv_cache);
-    DUMP_CACHE(conv1d_cache);
-    DUMP_CACHE(rglru_cache);
 #undef DUMP_CACHE
   }
   return sizeof(name) + sizeof(pos) + blob.total_size();
@@ -243,8 +224,6 @@ void load_impl(cgemma::session* sess, const char* buf, size_t n) {
     }                                                                                             \
   } while (false)
   LOAD_CACHE(kv_cache);
-  LOAD_CACHE(conv1d_cache);
-  LOAD_CACHE(rglru_cache);
 #undef LOAD_CACHE
 }