ggml : recurrent state rollback for ggml_ssm_scan (#26623)

* Initial changes for Recurrent state rollback for nemotron for cpu and cuda

* Removing CPU RS rollback. Will enable it in subsequent PRs

* addition of test case

* Removing assert and calling runtime API to check if op is supported

* removing extra API and updating the call sites for K

* replace static cuda detection to runtime fused_op api

* address review comments and fallback when SSM rollback not supprted

* Adding changes for supporting RS-rollback in CPU. Also added test-backend-ops for cpu and cuda

* removing memory manipulation as rs rollback is now supported in CPU

* removing the static probe which is not needed now

* correcting the format

* address review comments

* enabling test for all the backends, unsupported backends will fallback to CPU

* Apply suggestions from code review

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* choose different graph based on the result of fused_ssm_op is supported or not and also handled memory->n_rs_seq >1 case incase of op is not supported

* Support K > 1 in ssm_scan for all backends

* Fix CI Issues

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: Gaurav Garg <gaugarg@nvidia.com>

ggml/src/ggml-metal/ggml-metal-device.m

@@ -1376,9 +1376,10 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                 ggml_is_contiguous_rows(op->src[1]) &&
                 ggml_is_contiguous_rows(op->src[2]) &&
                 ggml_is_contiguous_rows(op->src[3]);
-        case GGML_OP_SSM_CONV:
         case GGML_OP_SSM_SCAN:
             return has_simdgroup_reduction;
+        case GGML_OP_SSM_CONV:
+            return has_simdgroup_reduction;
         case GGML_OP_RWKV_WKV6:
         case GGML_OP_RWKV_WKV7:
             return true;

ggml/src/ggml-metal/ggml-metal-impl.h

@@ -880,6 +880,7 @@ typedef struct {
     int64_t  n_group;
     int64_t  n_seq_tokens;
     int64_t  n_seqs;
+    int64_t  K;
     uint64_t s_off;
     uint64_t nb00;
     uint64_t nb01;

ggml/src/ggml-metal/ggml-metal-ops.cpp

@@ -1710,6 +1710,10 @@ int ggml_metal_op_ssm_scan(ggml_metal_op_t ctx, int idx) {
     const int64_t n_group      = ne41;
     const int64_t n_seq_tokens = ne12;
     const int64_t n_seqs       = ne13;
+    const int64_t K            = ggml_get_op_params_i32(op, 0);
+
+    GGML_ASSERT(K >= 1);
+    GGML_ASSERT(ggml_nelements(op->src[1]) + K*d_state*d_inner*n_head*n_seqs == ggml_nelements(op));
 
     ggml_metal_kargs_ssm_scan args = {
         /*.d_state      =*/ d_state,
@@ -1718,6 +1722,7 @@ int ggml_metal_op_ssm_scan(ggml_metal_op_t ctx, int idx) {
         /*.n_group      =*/ n_group,
         /*.n_seq_tokens =*/ n_seq_tokens,
         /*.n_seqs       =*/ n_seqs,
+        /*.K            =*/ K,
         /*.s_off        =*/ ggml_nelements(op->src[1]) * sizeof(float),
         /*.nb00         =*/ nb00,
         /*.nb01         =*/ nb01,

ggml/src/ggml-metal/ggml-metal.metal

@@ -2429,6 +2429,8 @@ kernel void kernel_ssm_scan_f32(
     const int32_t nh  = args.n_head;
     const int32_t ng  = args.n_group;
     const int32_t n_t = args.n_seq_tokens;
+    const int32_t n_s = args.n_seqs;
+    const int32_t K   = args.K;
 
     const int32_t s_off = args.s_off;
 
@@ -2487,6 +2489,12 @@ kernel void kernel_ssm_scan_f32(
             // recurse
             s0 = s;
 
+            const int32_t slot = n_t - 1 - (i2 + t);
+            if (slot > 0 && slot < K) {
+                device float * s_snapshot = (device float *) ((device char *) s_buff + (int64_t) slot*n_s*args.nb03);
+                s_snapshot[i] = s;
+            }
+
             B  += args.ns42;
             C  += args.ns52;
         }