llama : support Jamba hybrid Transformer-Mamba models (#7531)

* wip: llama : separate recurrent states from the KV cache

This will be necessary to support Jamba
(and other recurrent models mixed with Attention).

Doesn't compile yet, and finding a slot isn't yet done correctly for recurrent states.

* llama : use std::find for seq_nodes in llama_rs_cache

* llama : state checkpoints for recurrent models

* llama : correctly handle more edge cases for the rs cache

* llama : rename many llama_kv_cache_* functions

* llama : remove useless return value for some llama_cache_* functions

* llama : rethink recurrent state cell counts

* llama : begin work on support for variable GQA

This will also be useful for Jamba if we consider the Mamba layers
to have 0 KV heads.

* llama : gracefully fail when not finding hybrid slot

* llama : support Jamba

* llama : fix BERT inference without KV cache

* convert-hf : check for unprocessed Jamba experts

* convert-hf : support Mini-Jamba conversion

* llama : fix Jamba quantization sanity checks

* llama : sequence-length-aware batch splitting

* llama : use equal-sequence-length sub-batches for recurrent models

* ggml : simplify SSM-related operators

* llama : make recurrent state slot allocation contiguous

* llama : adapt internal uses of batches to llama_ubatch

* llama : fix batch split output count for embeddings

* llama : minimize swaps when reordering logits

This reduces overhead when running hellaswag
on thousands of sequences with very small 100k params Mamba models.

* llama : fix edge case finding batch seq_id of split recurrent cell

This otherwise was a problem when running the HellaSwag benchmark
with small batch sizes, making it crash.

* llama : avoid copies for simple batch splits

* ggml : make ggml_ssm_scan not modify its source tensors

* llama : fix shared recurrent tail cell count for small ubatch sizes

Otherwise it was impossible to run the 'parallel' example with '-ub 1'
with a Mamba or Jamba model.

* llama : fix .base() compilation error on Windows

* llama : allow doing the equivalent of SSM_CONV with SUM_ROWS and MUL

* ggml : allow GGML_OP_CONCAT to work on non-contiguous tensors

The implementation already supported it,
and this makes Mamba's conv step slightly faster.

* mamba : fix non-contiguous usage of ggml_silu

* llama : session saving and reloading for hybrid models

* convert_hf : fix Jamba conversion

* llama : fix mixed signedness comparison

* llama : use unused n_embd_k_gqa in k_shift

This also slightly reduces the diff from the master branch

* llama : begin renaming llama_past back to llama_kv_cache

* llama : remove implicit recurrent state rollbacks

* llama : partially apply clang-format style

* convert : fix jamba conv1d shape squeezing

* graph : add back hybrid memory graph input

But this time it contains the sub-cache graph inputs.
This *should* make it easier to handle updating the inputs
when caching the graph (eventually).

* model : add Jamba to Mamba-specific hparams printing

* jamba : remove redundant nullptr initializations

* model : remove unnecessary prefix for tensor loading constants

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* model : use ggml_swiglu_split for Mamba

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* model : make falcon-h1 use shared mamba2 layer builder

* memory : avoid referring to KV in recurrent cache logs

* gguf-py : avoid adding duplicate tensor mappings for Jamba

Some of the tensor names are common with Llama4

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

src/llama-memory-recurrent.cpp

@@ -25,9 +25,6 @@ llama_memory_recurrent::llama_memory_recurrent(
                  uint32_t    n_seq_max) : hparams(model.hparams), n_seq_max(n_seq_max) {
     const int32_t n_layer = hparams.n_layer;
 
-    LLAMA_LOG_INFO("%s: mem_size = %u, n_seq_max = %u, type_r = '%s', type_s = '%s', n_layer = %d\n",
-            __func__, mem_size, n_seq_max, ggml_type_name(type_r), ggml_type_name(type_s), n_layer);
-
     head = 0;
     size = mem_size;
     used = 0;
@@ -84,7 +81,7 @@ llama_memory_recurrent::llama_memory_recurrent(
 
         ggml_context * ctx = ctx_for_buft(buft);
         if (!ctx) {
-            throw std::runtime_error("failed to create ggml context for kv cache");
+            throw std::runtime_error("failed to create ggml context for rs cache");
         }
 
         ggml_tensor * r = ggml_new_tensor_1d(ctx, type_r, hparams.n_embd_r()*mem_size);
@@ -102,10 +99,10 @@ llama_memory_recurrent::llama_memory_recurrent(
 
         ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
         if (!buf) {
-            throw std::runtime_error("failed to allocate buffer for kv cache");
+            throw std::runtime_error("failed to allocate buffer for rs cache");
         }
         ggml_backend_buffer_clear(buf, 0);
-        LLAMA_LOG_INFO("%s: %10s KV buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0);
+        LLAMA_LOG_INFO("%s: %10s RS buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0);
         bufs.emplace_back(buf);
     }
 
@@ -113,8 +110,8 @@ llama_memory_recurrent::llama_memory_recurrent(
         const size_t memory_size_r = size_r_bytes();
         const size_t memory_size_s = size_s_bytes();
 
-        LLAMA_LOG_INFO("%s: KV self size  = %7.2f MiB, R (%s): %7.2f MiB, S (%s): %7.2f MiB\n", __func__,
-                (float)(memory_size_r + memory_size_s) / (1024.0f * 1024.0f),
+        LLAMA_LOG_INFO("%s: size = %7.2f MiB (%6u cells, %3d layers, %2u seqs), R (%s): %7.2f MiB, S (%s): %7.2f MiB\n", __func__,
+                (float)(memory_size_r + memory_size_s) / (1024.0f * 1024.0f), mem_size, n_layer, n_seq_max,
                 ggml_type_name(type_r), (float)memory_size_r / (1024.0f * 1024.0f),
                 ggml_type_name(type_s), (float)memory_size_s / (1024.0f * 1024.0f));
     }