llama : support Jamba hybrid Transformer-Mamba models

common/common.cpp

@@ -2541,7 +2541,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
         if (llama_model_has_decoder(model)) {
             llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
         }
-        llama_kv_cache_clear(lctx);
+        llama_past_clear(lctx);
         llama_synchronize(lctx);
         llama_reset_timings(lctx);
     }

convert_hf_to_gguf.py

@@ -2874,6 +2874,120 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
         return [(new_name, data_torch)]
 
 
+@Model.register("JambaForCausalLM")
+class JambaModel(Model):
+    model_arch = gguf.MODEL_ARCH.JAMBA
+
+    def get_vocab_base_pre(self, tokenizer) -> str:
+        del tokenizer  # unused
+
+        return "gpt-2"
+
+    def set_vocab(self):
+        if (self.dir_model / "tokenizer.model").is_file():
+            # Using Jamba's tokenizer.json causes errors on model load
+            # (something about "byte not found in vocab"),
+            # but there's a working tokenizer.model
+            self._set_vocab_sentencepiece()
+        else:
+            # Some Jamba models only have a tokenizer.json, which works.
+            self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        d_model = self.find_hparam(["hidden_size", "mamba_d_model"])
+        d_conv  = self.find_hparam(["mamba_d_conv"],  optional=True) or 4
+        d_inner = self.hparams["mamba_expand"] * d_model
+        d_state = self.find_hparam(["mamba_d_state"], optional=True) or 16
+        # ceiling division
+        # ref: https://stackoverflow.com/a/17511341/22827863
+        # ref: https://github.com/state-spaces/mamba/blob/ce59daea3a090d011d6476c6e5b97f6d58ddad8b/mamba_ssm/modules/mamba_simple.py#L58
+        dt_rank      = self.find_hparam(["mamba_dt_rank"], optional=True) or -(d_model // -16)
+        rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-6
+        n_kv_head = self.hparams["num_key_value_heads"]
+        attn_offset = self.hparams["attn_layer_offset"]
+        attn_period = self.hparams["attn_layer_period"]
+        n_kv_vec = [0 for _ in range(attn_offset)] + [
+            n_kv_head if (i - attn_offset) % attn_period == 0 else 0 for i in range(attn_offset, self.block_count)
+        ]
+
+        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_context_length(self.find_hparam(["max_position_embeddings", "n_ctx"]))
+        self.gguf_writer.add_embedding_length(d_model)
+        self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
+        self.gguf_writer.add_head_count(self.hparams["num_attention_heads"])
+        self.gguf_writer.add_head_count_kv(n_kv_vec)
+        self.gguf_writer.add_ssm_conv_kernel(d_conv)
+        self.gguf_writer.add_ssm_inner_size(d_inner)
+        self.gguf_writer.add_ssm_state_size(d_state)
+        self.gguf_writer.add_ssm_time_step_rank(dt_rank)
+        self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
+        self.gguf_writer.add_expert_count(self.hparams["num_experts"])
+        self.gguf_writer.add_expert_used_count(self.hparams["num_experts_per_tok"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+
+        # Mini-Jamba
+        name = name.replace(".moe.", ".feed_forward.")
+        if bid is not None:
+            moe_offset = self.hparams["expert_layer_offset"]
+            moe_period = self.hparams["expert_layer_period"]
+
+            if not (bid >= moe_offset and (bid - moe_offset) % moe_period == 0):
+                name = name.replace(".experts.0.", ".")
+
+        # process the experts separately
+        if ".feed_forward.experts." in name:
+            n_experts = self.hparams["num_experts"]
+
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+
+                # merge the experts into a single 3d tensor
+                for wid in ["down_proj", "gate_proj", "up_proj"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.feed_forward.experts.{xid}.{wid}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    # using the same merged name as qwen2moe
+                    merged_name = f"model.layers.{bid}.mlp.experts.{wid}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+
+                    yield new_name, data_torch
+            return
+
+        new_name = self.map_tensor_name(name)
+
+        if name.endswith(".A_log"):
+            logger.debug("A_log --> A ==> " + new_name)
+            data_torch = -torch.exp(data_torch)
+
+        yield new_name, data_torch
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+
 @Model.register("CohereForCausalLM")
 class CommandR2Model(Model):
     model_arch = gguf.MODEL_ARCH.COMMAND_R

examples/batched-bench/batched-bench.cpp

@@ -153,7 +153,7 @@ int main(int argc, char ** argv) {
 
                 const auto t_pp_start = ggml_time_us();
 
-                llama_kv_cache_clear(ctx);
+                llama_past_clear(ctx);
 
                 if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
                     LOG_TEE("%s: llama_decode() failed\n", __func__);
@@ -162,7 +162,7 @@ int main(int argc, char ** argv) {
 
                 if (is_pp_shared) {
                     for (int32_t i = 1; i < pl; ++i) {
-                        llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+                        llama_past_seq_cp(ctx, 0, i, -1, -1);
                     }
                 }
 

examples/batched.swift/Sources/main.swift

@@ -98,7 +98,7 @@ if llama_decode(context, batch) != 0 {
 }
 
 for i in 1 ..< n_parallel {
-    llama_kv_cache_seq_cp(context, 0, Int32(i), 0, batch.n_tokens)
+    llama_past_seq_cp(context, 0, Int32(i), -1, -1)
 }
 
 if n_parallel > 1 {

examples/batched/batched.cpp

@@ -132,7 +132,7 @@ int main(int argc, char ** argv) {
     //// assign the system KV cache to all parallel sequences
     //// this way, the parallel sequences will "reuse" the prompt tokens without having to copy them
     //for (int32_t i = 1; i < n_parallel; ++i) {
-    //    llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+    //    llama_past_seq_cp(ctx, 0, i, -1, -1);
     //}
 
     if (n_parallel > 1) {

examples/cvector-generator/cvector-generator.cpp

@@ -338,7 +338,7 @@ static bool cb_eval(struct ggml_tensor * t, bool ask, void * user_data) {
 }
 
 static bool get_hidden_layers(llama_context * ctx, std::vector<llama_token> & tokens) {
-    llama_kv_cache_clear(ctx);
+    llama_past_clear(ctx);
     if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
         fprintf(stderr, "%s : failed to eval\n", __func__);
         return false;

examples/embedding/embedding.cpp

@@ -35,7 +35,7 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
     const struct llama_model * model = llama_get_model(ctx);
 
     // clear previous kv_cache values (irrelevant for embeddings)
-    llama_kv_cache_clear(ctx);
+    llama_past_clear(ctx);
 
     // run model
     fprintf(stderr, "%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);

examples/gritlm/gritlm.cpp

@@ -43,7 +43,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
         }
 
         // clear previous kv_cache values (irrelevant for embeddings)
-        llama_kv_cache_clear(ctx);
+        llama_past_clear(ctx);
         llama_set_embeddings(ctx, true);
         llama_set_causal_attn(ctx, false);
 
@@ -98,7 +98,7 @@ static std::string generate(llama_context * ctx, const std::string & prompt, boo
     const llama_model * mdl = llama_get_model(ctx);
     llama_token eos_token = llama_token_eos(mdl);
 
-    llama_kv_cache_clear(ctx);
+    llama_past_clear(ctx);
     llama_set_embeddings(ctx, false);
     llama_set_causal_attn(ctx, true);
 

examples/imatrix/imatrix.cpp

@@ -499,7 +499,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
         const auto t_start = std::chrono::high_resolution_clock::now();
 
         // clear the KV cache
-        llama_kv_cache_clear(ctx);
+        llama_past_clear(ctx);
 
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;

examples/infill/infill.cpp

@@ -385,8 +385,8 @@ int main(int argc, char ** argv) {
                 LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
                     n_past, n_left, n_ctx, params.n_keep, n_discard);
 
-                llama_kv_cache_seq_rm (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
-                llama_kv_cache_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+                llama_past_seq_rm (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+                llama_past_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
 
                 n_past -= n_discard;
 

examples/llama-bench/llama-bench.cpp

@@ -1515,7 +1515,7 @@ int main(int argc, char ** argv) {
 
         test t(inst, lmodel, ctx);
 
-        llama_kv_cache_clear(ctx);
+        llama_past_clear(ctx);
 
         // cool off before the test
         if (params.delay) {
@@ -1549,7 +1549,7 @@ int main(int argc, char ** argv) {
         }
 
         for (int i = 0; i < params.reps; i++) {
-            llama_kv_cache_clear(ctx);
+            llama_past_clear(ctx);
 
             uint64_t t_start = get_time_ns();
 

examples/llama.android/llama/src/main/cpp/llama-android.cpp

@@ -194,7 +194,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
         }
 
         batch->logits[batch->n_tokens - 1] = true;
-        llama_kv_cache_clear(context);
+        llama_past_clear(context);
 
         const auto t_pp_start = ggml_time_us();
         if (llama_decode(context, *batch) != 0) {
@@ -206,7 +206,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
 
         LOGi("Benchmark text generation (tg)");
 
-        llama_kv_cache_clear(context);
+        llama_past_clear(context);
         const auto t_tg_start = ggml_time_us();
         for (i = 0; i < tg; i++) {
 
@@ -223,7 +223,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
 
         const auto t_tg_end = ggml_time_us();
 
-        llama_kv_cache_clear(context);
+        llama_past_clear(context);
 
         const auto t_pp = double(t_pp_end - t_pp_start) / 1000000.0;
         const auto t_tg = double(t_tg_end - t_tg_start) / 1000000.0;
@@ -439,5 +439,5 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop(
 extern "C"
 JNIEXPORT void JNICALL
 Java_android_llama_cpp_LLamaAndroid_kv_1cache_1clear(JNIEnv *, jobject, jlong context) {
-    llama_kv_cache_clear(reinterpret_cast<llama_context *>(context));
+    llama_past_clear(reinterpret_cast<llama_context *>(context));
 }

examples/llama.swiftui/llama.cpp.swift/LibLlama.swift

@@ -216,7 +216,7 @@ actor LlamaContext {
             }
             batch.logits[Int(batch.n_tokens) - 1] = 1 // true
 
-            llama_kv_cache_clear(context)
+            llama_past_clear(context)
 
             let t_pp_start = ggml_time_us()
 
@@ -229,7 +229,7 @@ actor LlamaContext {
 
             // bench text generation
 
-            llama_kv_cache_clear(context)
+            llama_past_clear(context)
 
             let t_tg_start = ggml_time_us()
 
@@ -248,7 +248,7 @@ actor LlamaContext {
 
             let t_tg_end = ggml_time_us()
 
-            llama_kv_cache_clear(context)
+            llama_past_clear(context)
 
             let t_pp = Double(t_pp_end - t_pp_start) / 1000000.0
             let t_tg = Double(t_tg_end - t_tg_start) / 1000000.0
@@ -298,7 +298,7 @@ actor LlamaContext {
     func clear() {
         tokens_list.removeAll()
         temporary_invalid_cchars.removeAll()
-        llama_kv_cache_clear(context)
+        llama_past_clear(context)
     }
 
     private func tokenize(text: String, add_bos: Bool) -> [llama_token] {

examples/lookahead/lookahead.cpp

@@ -96,7 +96,7 @@ int main(int argc, char ** argv) {
     llama_decode(ctx, llama_batch_get_one(&inp.back(),           1, n_input - 1, 0));
 
     for (int s = 1; s < W + G + 1; ++s) {
-        llama_kv_cache_seq_cp(ctx, 0, s, -1, -1);
+        llama_past_seq_cp(ctx, 0, s, -1, -1);
     }
 
     const auto t_enc_end = ggml_time_us();
@@ -438,17 +438,18 @@ int main(int argc, char ** argv) {
 
         // KV cache management
         // if no verification token matched, we simply remove all cells from this batch -> no fragmentation
-        llama_kv_cache_seq_rm(ctx, -1, n_past, -1);
+        // FIXME: recurrent and hybrid models
+        llama_past_seq_rm(ctx, -1, n_past, -1);
 
         if (seq_id_best != 0) {
             // if a verification token matched, we keep the best sequence and remove the rest
             // this leads to some KV cache fragmentation
-            llama_kv_cache_seq_keep(ctx, seq_id_best);
-            llama_kv_cache_seq_cp  (ctx, seq_id_best, 0, -1, -1);
-            llama_kv_cache_seq_rm  (ctx, seq_id_best,    -1, -1);
+            llama_past_seq_keep(ctx, seq_id_best);
+            llama_past_seq_cp  (ctx, seq_id_best, 0, -1, -1);
+            llama_past_seq_rm  (ctx, seq_id_best,    -1, -1);
 
             for (int s = 1; s < W + G + 1; ++s) {
-                llama_kv_cache_seq_cp(ctx, 0, s, -1, -1);
+                llama_past_seq_cp(ctx, 0, s, -1, -1);
             }
         }
     }

examples/lookup/lookup.cpp

@@ -194,7 +194,8 @@ int main(int argc, char ** argv){
 
         // KV cache management
         // clean the cache of draft tokens that weren't accepted
-        llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
+        // FIXME: recurrent and hybrid models
+        llama_past_seq_rm(ctx, 0, n_past, -1);
 
         llama_batch_clear(batch_tgt);
         llama_batch_add(batch_tgt, draft[0], n_past, { 0 }, true);