tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-08-05 00:25:26 -04:00
Code Issues Packages Projects Releases Wiki Activity

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

Browse Source 
* 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>
This commit is contained in:
compilade
2025-07-09 14:59:57 -04:00
committed by GitHub
parent 98bab638fb
commit 4a5686da22
10 changed files with 622 additions and 423 deletions
Download Patch File Download Diff File Expand all files Collapse all files

58
src/llama-graph.h
View File

@@ -322,32 +322,21 @@ public:
class llm_graph_input_mem_hybrid : public llm_graph_input_i {
public:
llm_graph_input_mem_hybrid(
const llama_hparams & hparams,
const llama_cparams & cparams,
const llama_memory_hybrid_context * mctx) :
hparams(hparams),
cparams(cparams),
mctx(mctx) {
}
std::unique_ptr<llm_graph_input_attn_kv_unified> inp_attn,
std::unique_ptr<llm_graph_input_rs> inp_rs,
const llama_memory_hybrid_context * mctx) :
inp_attn(std::move(inp_attn)),
inp_rs(std::move(inp_rs)),
mctx(mctx) { }
virtual ~llm_graph_input_mem_hybrid() = default;
void set_input(const llama_ubatch * ubatch) override;
ggml_tensor * s_copy; // I32 [kv_size]
std::unique_ptr<llm_graph_input_attn_kv_unified> inp_attn;
std::unique_ptr<llm_graph_input_rs> inp_rs;
ggml_tensor * get_k_idxs() const { return self_k_idxs; }
ggml_tensor * get_v_idxs() const { return self_v_idxs; }
ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; }
ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]
ggml_tensor * self_v_idxs = nullptr; // I64 [n_batch]
ggml_tensor * self_kq_mask = nullptr; // F32 [n_kv, n_batch, 1, 1]
ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch, 1, 1]
const llama_hparams & hparams;
const llama_cparams & cparams;
llm_graph_input_attn_kv_unified * get_attn() const { return inp_attn.get(); }
llm_graph_input_rs * get_recr() const { return inp_rs.get(); }
const llama_memory_hybrid_context * mctx;
};
@@ -579,8 +568,6 @@ struct llm_graph_context {
ggml_tensor * build_inp_pos_bucket_dec() const;
ggml_tensor * build_pos_bias(ggml_tensor * pos_bucket, ggml_tensor * attn_rel_b) const;
llm_graph_input_mem_hybrid * build_inp_mem_hybrid() const;
//
// attention
//
@@ -656,18 +643,6 @@ struct llm_graph_context {
float kq_scale,
int il) const;
ggml_tensor * build_attn(
llm_graph_input_mem_hybrid * inp,
ggml_cgraph * gf,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
ggml_tensor * kq_b,
ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
float kq_scale,
int il) const;
//
// recurrent
//
@@ -700,14 +675,6 @@ struct llm_graph_context {
int32_t n_seqs,
const llm_graph_get_rows_fn & get_state_rows = ggml_get_rows) const;
ggml_tensor * build_rs(
llm_graph_input_mem_hybrid * inp,
ggml_cgraph * gf,
ggml_tensor * s,
int32_t state_size,
int32_t n_seqs,
const llm_graph_get_rows_fn & get_state_rows = ggml_get_rows) const;
ggml_tensor * build_rwkv_token_shift_load(
llm_graph_input_rs * inp,
ggml_cgraph * gf,
@@ -718,6 +685,11 @@ struct llm_graph_context {
ggml_tensor * token_shift,
const llama_ubatch & ubatch,
int il) const;
//
// hybrid
//
llm_graph_input_mem_hybrid * build_inp_mem_hybrid() const;
//
// pooling