metal: SSM kernel improvements #17876

gabe-l-hart · 2025-12-08T23:52:03Z

Description

This branch adds performance improvements on metal for both SSM_CONV and SSM_SCAN. The kernels were heavily edited by Claude Code, but I've reviewed all changes.

Changes

SSM_CONV: Implement a batched version that uses batches of 256 threads for multi-token prefill.
- Split what was the y dim of the outer grid into y / BATCH_SIZE and use BATCH_SIZE as x for threadgroup (inner grid)
- Recompute offsets from tgpig.y and tpitg.x
SSM_SCAN: Reduce redundant x_dt / dA computations

Performance

./bin/llama-batched-bench -m ~/models/ibm-granite/granite-4.0-h-1b/ggml-model-Q8_0.gguf -c 131072 -b 2048 -ub 512 -npp 1024,4096,8192 -ntg 128 -npl 1,4,8 -ngl 99

Baseline (`c8554b6`)

PP	TG	B	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s	T s	S t/s
1024	128	1	1152	0.719	1423.52	1.407	91.00	2.126	541.87
1024	128	4	4608	2.877	1423.62	2.947	173.72	5.824	791.15
1024	128	8	9216	5.748	1425.13	5.269	194.33	11.018	836.48
4096	128	1	4224	2.887	1418.90	1.415	90.49	4.301	982.03
4096	128	4	16896	11.537	1420.09	2.990	171.24	14.527	1163.05
4096	128	8	33792	23.169	1414.28	6.213	164.80	29.383	1150.06
8192	128	1	8320	6.458	1268.54	1.437	89.11	7.894	1053.92
8192	128	4	33280	23.539	1392.09	3.236	158.24	26.774	1242.98
8192	128	8	66560	47.488	1380.06	5.973	171.43	53.461	1245.02

SSM_CONV improvements

PP	TG	B	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s	T s	S t/s
1024	128	1	1152	0.457	2240.26	1.407	91.00	1.864	618.12
1024	128	4	4608	1.821	2249.29	2.948	173.68	4.769	966.24
1024	128	8	9216	3.641	2250.21	5.265	194.47	8.906	1034.80
4096	128	1	4224	1.834	2233.20	1.410	90.76	3.244	1301.91
4096	128	4	16896	7.332	2234.56	2.997	170.86	10.329	1635.82
4096	128	8	33792	14.683	2231.72	5.347	191.51	20.030	1687.09
8192	128	1	8320	3.723	2200.12	1.425	89.81	5.149	1615.96
8192	128	4	33280	15.000	2184.60	3.074	166.54	18.074	1841.32
8192	128	8	66560	33.971	1929.15	6.071	168.67	40.043	1662.23

SSM_CONV + SSM_SCAN improvements

PP	TG	B	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s	T s	S t/s
1024	128	1	1152	0.437	2345.00	1.411	90.75	1.847	623.65
1024	128	4	4608	1.732	2364.53	2.977	171.98	4.709	978.47
1024	128	8	9216	3.487	2349.39	5.374	190.53	8.861	1040.03
4096	128	1	4224	1.753	2336.90	1.425	89.80	3.178	1329.05
4096	128	4	16896	7.007	2338.09	3.020	169.55	10.027	1685.03
4096	128	8	33792	14.042	2333.49	5.412	189.22	19.454	1737.01
8192	128	1	8320	3.572	2293.39	1.434	89.25	5.006	1661.94
8192	128	4	33280	14.758	2220.40	3.208	159.62	17.965	1852.45
8192	128	8	66560	33.036	1983.79	6.053	169.17	39.089	1702.79

jeffbolznv · 2025-12-09T05:47:12Z

Mind adding some representative cases to test-backend-ops perf?

ggerganov · 2025-12-09T11:20:33Z

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

+    } else {
+        auto pipeline = ggml_metal_library_get_pipeline_ssm_conv(lib, op);
+
+        ggml_metal_encoder_set_pipeline(enc, pipeline);
+        ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
+        ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+        ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[1]), 2);
+        ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op),         3);
+
+        ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne1, ne02, 1, 1, 1);
+    }


Is the old kernel faster for ne1 == 1? If not, we can remove it and always use the batched kernel?

Good question, I'll test that today.

It looks like non-batched is significantly faster for ne1 == 1, so I think we should keep both paths.

ggerganov · 2025-12-09T12:30:27Z

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

+    // Use batched kernel for prefill (ne1 > 1) to reduce threadgroup dispatch overhead
+    constexpr int BATCH_SIZE = 256;
+    const bool use_batched = (ne1 > 1);

-    ggml_metal_encoder_set_pipeline(enc, pipeline);
-    ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
-    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 1);
-    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[1]), 2);
-    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op),         3);
+    if (use_batched) {
+        auto pipeline = ggml_metal_library_get_pipeline_ssm_conv_batched(lib, op);
+
+        ggml_metal_encoder_set_pipeline(enc, pipeline);
+        ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
+        ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+        ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[1]), 2);
+        ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op),         3);

-    ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne1, ne02, 1, 1, 1);
+        // Dispatch: ne01 rows, ceil(ne1/BATCH_SIZE) token batches, ne02 sequences
+        // Each threadgroup has BATCH_SIZE threads, each handling one token
+        const int n_token_batches = (ne1 + BATCH_SIZE - 1) / BATCH_SIZE;
+        ggml_metal_encoder_dispatch_threadgroups(enc, ne01, n_token_batches, ne02, BATCH_SIZE, 1, 1);
+    } else {
+        auto pipeline = ggml_metal_library_get_pipeline_ssm_conv(lib, op);
+
+        ggml_metal_encoder_set_pipeline(enc, pipeline);
+        ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
+        ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+        ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[1]), 2);
+        ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op),         3);
+
+        ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne1, ne02, 1, 1, 1);
+    }


Interesting - I am quite surprised that this change makes such a big difference. I have to try this approach for all other kernels that launch threadgroups with just 1 thread: unary, binary, scale, clamp, fill, etc.

Honestly, I was very surprised too. All credit to Claude Code with Opus 4.5 for the insight.

ggerganov · 2025-12-09T13:11:10Z

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


+// Batched version: each threadgroup processes multiple tokens for better efficiency
+// Thread layout: each thread handles one token, threadgroup covers BATCH_SIZE tokens
+template<int BATCH_SIZE>


This parameter does not have to be a template. The better pattern is to make it a function constant. For example, see how this works:

llama.cpp/ggml/src/ggml-metal/ggml-metal.metal

Lines 3108 to 3122 in 86a3f0f

constant short FC_mul_mv_nsg [[function_constant(FC_MUL_MV + 0)]];

constant short FC_mul_mv_nxpsg [[function_constant(FC_MUL_MV + 1)]];

template<typename block_q_type, short NR0, typename args_t>

void mul_vec_q_n_f32_impl(

args_t args,

device const char * src0,

device const char * src1,

device char * dst,

threadgroup char * shmem,

uint3 tgpig,

ushort tiisg,

ushort sgitg) {

const short NSG = FC_mul_mv_nsg;

Then observe how we pass the function constants during the construction of the pipeline:

llama.cpp/ggml/src/ggml-metal/ggml-metal-device.cpp

Lines 472 to 494 in 86a3f0f

ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv_ext(ggml_metal_library_t lib, ggml_type tsrc0, ggml_type tsrc1, int nsg, int nxpsg, int r1ptg) {

char base[256];

char name[256];

snprintf(base, 256, "kernel_mul_mv_ext_%s_%s_r1_%d", ggml_type_name(tsrc0), ggml_type_name(tsrc1), r1ptg);

snprintf(name, 256, "%s_nsg=%d_nxpsg=%d", base, nsg, nxpsg);

ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);

if (!res.pipeline) {

ggml_metal_cv_t cv = ggml_metal_cv_init();

ggml_metal_cv_set_int16(cv, nsg, FC_MUL_MV + 0);

ggml_metal_cv_set_int16(cv, nxpsg, FC_MUL_MV + 1);

res = ggml_metal_library_compile_pipeline(lib, base, name, cv);

ggml_metal_cv_free(cv);

}

return res;

}

This way, we can construct pipelines at runtime with different FC_ssm_conv_bs, instead of defining many template instantiations that would increase the compile time.

When you do that, add logic in ggml_metal_op_ssm_conv to determine the smallest power of 2 that is larger or equal to ne1 and less or equal than 256. Use that power to construct a pipeline with the respective batch size.

For example, if ne1 == 100, we want a pipeline with FC_ssm_conv_bs == 128. And so on.

gabe-l-hart · 2025-12-09T16:44:52Z

Test results for `test-backend-ops perf -o SSM_CONV -b Metal`

@jeffbolznv I've honestly never quite known how to interpret the results of test-backend-ops perf. Intuitively, I would think a higher us/run would mean slower and a higher GB/s would mean more efficient throughput, but these results seem to show I have that exactly backwards.

@ggerganov Assuming my intuition is backwards, then it looks like the batched implementation is indeed always faster. Would it be worth also adding a float4 variant of the batch implementation?

Test results for test-backend-ops perf:

Baseline (non-batch)

ggml_metal_library_compile_pipeline: compiling pipeline: base = 'kernel_ssm_conv_f32_f32', name = 'kernel_ssm_conv_f32_f32'
ggml_metal_library_compile_pipeline: loaded kernel_ssm_conv_f32_f32                       0x12010a0a0 | th_max = 1024 | th_width =   32
  SSM_CONV(type=f32,ne_a=[4,1024,1,1],ne_b=[3,1024,1,1]):             139264 runs -     7.19 us/run -       36 kB/run -    4.77 GB/s
  SSM_CONV(type=f32,ne_a=[8,1024,1,1],ne_b=[3,1024,1,1]):              65536 runs -    16.12 us/run -       68 kB/run -    4.02 GB/s
  SSM_CONV(type=f32,ne_a=[4,1024,4,1],ne_b=[3,1024,1,1]):              49152 runs -    20.51 us/run -      108 kB/run -    5.02 GB/s
  SSM_CONV(type=f32,ne_a=[4,1536,1,1],ne_b=[3,1536,1,1]):             114688 runs -     9.31 us/run -       54 kB/run -    5.53 GB/s
  SSM_CONV(type=f32,ne_a=[8,1536,1,1],ne_b=[3,1536,1,1]):              49152 runs -    22.91 us/run -      102 kB/run -    4.25 GB/s
  SSM_CONV(type=f32,ne_a=[4,1536,4,1],ne_b=[3,1536,1,1]):              40960 runs -    29.60 us/run -      162 kB/run -    5.22 GB/s
  SSM_CONV(type=f32,ne_a=[4,2048,1,1],ne_b=[3,2048,1,1]):              90112 runs -    11.61 us/run -       72 kB/run -    5.92 GB/s
  SSM_CONV(type=f32,ne_a=[8,2048,1,1],ne_b=[3,2048,1,1]):              40960 runs -    29.68 us/run -      136 kB/run -    4.37 GB/s
  SSM_CONV(type=f32,ne_a=[4,2048,4,1],ne_b=[3,2048,1,1]):              32768 runs -    38.45 us/run -      216 kB/run -    5.36 GB/s
ggml_metal_library_compile_pipeline: compiling pipeline: base = 'kernel_ssm_conv_f32_f32_4', name = 'kernel_ssm_conv_f32_f32_4'
ggml_metal_library_compile_pipeline: loaded kernel_ssm_conv_f32_f32_4                     0x11df051c0 | th_max = 1024 | th_width =   32
  SSM_CONV(type=f32,ne_a=[4,1024,1,1],ne_b=[4,1024,1,1]):             204800 runs -     5.04 us/run -       36 kB/run -    6.82 GB/s
  SSM_CONV(type=f32,ne_a=[8,1024,1,1],ne_b=[4,1024,1,1]):              73728 runs -    14.26 us/run -       68 kB/run -    4.55 GB/s
  SSM_CONV(type=f32,ne_a=[4,1024,4,1],ne_b=[4,1024,1,1]):              90112 runs -    11.88 us/run -       96 kB/run -    7.71 GB/s
  SSM_CONV(type=f32,ne_a=[4,1536,1,1],ne_b=[4,1536,1,1]):             163840 runs -     6.13 us/run -       54 kB/run -    8.40 GB/s
  SSM_CONV(type=f32,ne_a=[8,1536,1,1],ne_b=[4,1536,1,1]):              57344 runs -    19.96 us/run -      102 kB/run -    4.87 GB/s
  SSM_CONV(type=f32,ne_a=[4,1536,4,1],ne_b=[4,1536,1,1]):              65536 runs -    16.36 us/run -      144 kB/run -    8.39 GB/s
  SSM_CONV(type=f32,ne_a=[4,2048,1,1],ne_b=[4,2048,1,1]):             139264 runs -     7.38 us/run -       72 kB/run -    9.30 GB/s
  SSM_CONV(type=f32,ne_a=[8,2048,1,1],ne_b=[4,2048,1,1]):              40960 runs -    25.68 us/run -      136 kB/run -    5.05 GB/s
  SSM_CONV(type=f32,ne_a=[4,2048,4,1],ne_b=[4,2048,1,1]):              49152 runs -    20.84 us/run -      192 kB/run -    8.78 GB/s

Batched for n_t > 1

ggml_metal_library_compile_pipeline: compiling pipeline: base = 'kernel_ssm_conv_f32_f32_b256', name = 'kernel_ssm_conv_f32_f32_b256'
ggml_metal_library_compile_pipeline: loaded kernel_ssm_conv_f32_f32_b256                  0x107505090 | th_max = 1024 | th_width =   32
  SSM_CONV(type=f32,ne_a=[4,1024,1,1],ne_b=[3,1024,1,1]):              57344 runs -    19.91 us/run -       36 kB/run -    1.72 GB/s
  SSM_CONV(type=f32,ne_a=[8,1024,1,1],ne_b=[3,1024,1,1]):              57344 runs -    20.04 us/run -       68 kB/run -    3.24 GB/s
  SSM_CONV(type=f32,ne_a=[4,1024,4,1],ne_b=[3,1024,1,1]):              49152 runs -    21.20 us/run -      108 kB/run -    4.86 GB/s
  SSM_CONV(type=f32,ne_a=[4,1536,1,1],ne_b=[3,1536,1,1]):              40960 runs -    28.42 us/run -       54 kB/run -    1.81 GB/s
  SSM_CONV(type=f32,ne_a=[8,1536,1,1],ne_b=[3,1536,1,1]):              40960 runs -    28.68 us/run -      102 kB/run -    3.39 GB/s
  SSM_CONV(type=f32,ne_a=[4,1536,4,1],ne_b=[3,1536,1,1]):              32768 runs -    30.60 us/run -      162 kB/run -    5.05 GB/s
  SSM_CONV(type=f32,ne_a=[4,2048,1,1],ne_b=[3,2048,1,1]):              32768 runs -    37.49 us/run -       72 kB/run -    1.83 GB/s
  SSM_CONV(type=f32,ne_a=[8,2048,1,1],ne_b=[3,2048,1,1]):              32768 runs -    37.75 us/run -      136 kB/run -    3.44 GB/s
  SSM_CONV(type=f32,ne_a=[4,2048,4,1],ne_b=[3,2048,1,1]):              32768 runs -    40.05 us/run -      216 kB/run -    5.14 GB/s
ggml_metal_library_compile_pipeline: compiling pipeline: base = 'kernel_ssm_conv_f32_f32_4', name = 'kernel_ssm_conv_f32_f32_4'
ggml_metal_library_compile_pipeline: loaded kernel_ssm_conv_f32_f32_4                     0x301505560 | th_max = 1024 | th_width =   32
  SSM_CONV(type=f32,ne_a=[4,1024,1,1],ne_b=[4,1024,1,1]):             204800 runs -     5.03 us/run -       36 kB/run -    6.82 GB/s
  SSM_CONV(type=f32,ne_a=[8,1024,1,1],ne_b=[4,1024,1,1]):              57344 runs -    19.56 us/run -       68 kB/run -    3.31 GB/s
  SSM_CONV(type=f32,ne_a=[4,1024,4,1],ne_b=[4,1024,1,1]):              90112 runs -    11.93 us/run -       96 kB/run -    7.68 GB/s
  SSM_CONV(type=f32,ne_a=[4,1536,1,1],ne_b=[4,1536,1,1]):             163840 runs -     6.17 us/run -       54 kB/run -    8.35 GB/s
  SSM_CONV(type=f32,ne_a=[8,1536,1,1],ne_b=[4,1536,1,1]):              40960 runs -    28.45 us/run -      102 kB/run -    3.42 GB/s
  SSM_CONV(type=f32,ne_a=[4,1536,4,1],ne_b=[4,1536,1,1]):              65536 runs -    16.32 us/run -      144 kB/run -    8.41 GB/s
  SSM_CONV(type=f32,ne_a=[4,2048,1,1],ne_b=[4,2048,1,1]):             139264 runs -     7.36 us/run -       72 kB/run -    9.33 GB/s
  SSM_CONV(type=f32,ne_a=[8,2048,1,1],ne_b=[4,2048,1,1]):              32768 runs -    37.16 us/run -      136 kB/run -    3.49 GB/s
  SSM_CONV(type=f32,ne_a=[4,2048,4,1],ne_b=[4,2048,1,1]):              49152 runs -    20.85 us/run -      192 kB/run -    8.78 GB/s

Batched always

  SSM_CONV(type=f32,ne_a=[4,1024,1,1],ne_b=[3,1024,1,1]):              57344 runs -    19.91 us/run -       36 kB/run -    1.72 GB/s
  SSM_CONV(type=f32,ne_a=[8,1024,1,1],ne_b=[3,1024,1,1]):              57344 runs -    19.81 us/run -       68 kB/run -    3.27 GB/s
  SSM_CONV(type=f32,ne_a=[4,1024,4,1],ne_b=[3,1024,1,1]):              49152 runs -    21.16 us/run -      108 kB/run -    4.87 GB/s
  SSM_CONV(type=f32,ne_a=[4,1536,1,1],ne_b=[3,1536,1,1]):              40960 runs -    28.41 us/run -       54 kB/run -    1.81 GB/s
  SSM_CONV(type=f32,ne_a=[8,1536,1,1],ne_b=[3,1536,1,1]):              40960 runs -    28.69 us/run -      102 kB/run -    3.39 GB/s
  SSM_CONV(type=f32,ne_a=[4,1536,4,1],ne_b=[3,1536,1,1]):              32768 runs -    30.57 us/run -      162 kB/run -    5.05 GB/s
  SSM_CONV(type=f32,ne_a=[4,2048,1,1],ne_b=[3,2048,1,1]):              32768 runs -    37.20 us/run -       72 kB/run -    1.85 GB/s
  SSM_CONV(type=f32,ne_a=[8,2048,1,1],ne_b=[3,2048,1,1]):              32768 runs -    37.75 us/run -      136 kB/run -    3.44 GB/s
  SSM_CONV(type=f32,ne_a=[4,2048,4,1],ne_b=[3,2048,1,1]):              32768 runs -    39.89 us/run -      216 kB/run -    5.16 GB/s
  # NOTE: n_t == 1
  SSM_CONV(type=f32,ne_a=[4,1024,1,1],ne_b=[4,1024,1,1]):              57344 runs -    19.49 us/run -       36 kB/run -    1.76 GB/s
  SSM_CONV(type=f32,ne_a=[8,1024,1,1],ne_b=[4,1024,1,1]):              57344 runs -    19.60 us/run -       68 kB/run -    3.31 GB/s
  SSM_CONV(type=f32,ne_a=[4,1024,4,1],ne_b=[4,1024,1,1]):              49152 runs -    20.92 us/run -       96 kB/run -    4.38 GB/s
  # NOTE: n_t == 1
  SSM_CONV(type=f32,ne_a=[4,1536,1,1],ne_b=[4,1536,1,1]):              40960 runs -    28.35 us/run -       54 kB/run -    1.82 GB/s
  SSM_CONV(type=f32,ne_a=[8,1536,1,1],ne_b=[4,1536,1,1]):              40960 runs -    28.43 us/run -      102 kB/run -    3.42 GB/s
  SSM_CONV(type=f32,ne_a=[4,1536,4,1],ne_b=[4,1536,1,1]):              40960 runs -    29.87 us/run -      144 kB/run -    4.60 GB/s
  # NOTE: n_t == 1
  SSM_CONV(type=f32,ne_a=[4,2048,1,1],ne_b=[4,2048,1,1]):              32768 runs -    36.92 us/run -       72 kB/run -    1.86 GB/s
  SSM_CONV(type=f32,ne_a=[8,2048,1,1],ne_b=[4,2048,1,1]):              32768 runs -    37.20 us/run -      136 kB/run -    3.49 GB/s
  SSM_CONV(type=f32,ne_a=[4,2048,4,1],ne_b=[4,2048,1,1]):              32768 runs -    38.95 us/run -      192 kB/run -    4.70 GB/s

jeffbolznv · 2025-12-09T16:50:17Z

You do want lower us/run, higher GB/s. The two values are the same data, just GB/s is computed by summing tensor sizes and dividing by the runtime.

ggerganov · 2025-12-09T16:52:32Z

@gabe-l-hart These perf numbers don't add up to the observed llama-batched-bench results. Ideally, note down the actual shapes that are used during prompt processing and text generation and add tests that correspond to those shapes.

gabe-l-hart · 2025-12-09T16:53:35Z

Ok, I'm glad I'm not crazy! This does seem very fishy given the improved results with pp. I'll try to make these a closer representative test.

This was done using Claude Code. It found a number of optimizations around how the threads were organized, resulting in a huge performance boost! Branch: Mamba2SSD Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

This used Claude Code and resulted in a modest performance improvement while maintaining correctness. Branch: Mamba2SSD Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

Branch: SSMKernelImprovements Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

gabe-l-hart · 2025-12-09T17:17:15Z

Oh boy, that makes much more sense!

    // Examples from granite-4.0-h-1b/ggml-model-Q8_0.gguf
    // d_inner == 3072
    // d_conv == 4
    test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {515, 3328, 1, 1}, {4, 3328, 1, 1}));

Batched

ggml_metal_library_compile_pipeline: compiling pipeline: base = 'kernel_ssm_conv_f32_f32_b256', name = 'kernel_ssm_conv_f32_f32_b256'
ggml_metal_library_compile_pipeline: loaded kernel_ssm_conv_f32_f32_b256                  0x136b083d0 | th_max = 1024 | th_width =   32
  SSM_CONV(type=f32,ne_a=[515,3328,1,1],ne_b=[4,3328,1,1]):                    15024 runs -    70.64 us/run -    13403 kB/run -  180.96 GB/s

Un-Batched

ggml_metal_library_compile_pipeline: compiling pipeline: base = 'kernel_ssm_conv_f32_f32_4', name = 'kernel_ssm_conv_f32_f32_4'
ggml_metal_library_compile_pipeline: loaded kernel_ssm_conv_f32_f32_4                     0x12320be70 | th_max = 1024 | th_width =   32
  SSM_CONV(type=f32,ne_a=[515,3328,1,1],ne_b=[4,3328,1,1]):                     2504 runs -  3801.15 us/run -    13403 kB/run -    3.36 GB/s

gabe-l-hart · 2025-12-09T17:24:20Z

With a single-token (generate) example:

Batch for prefill only

ggml_metal_library_compile_pipeline: compiling pipeline: base = 'kernel_ssm_conv_f32_f32_b256', name = 'kernel_ssm_conv_f32_f32_b256'
ggml_metal_library_compile_pipeline: loaded kernel_ssm_conv_f32_f32_b256                  0x1292076d0 | th_max = 1024 | th_width =   32
  SSM_CONV(type=f32,ne_a=[515,3328,1,1],ne_b=[4,3328,1,1]):                    15024 runs -    70.69 us/run -    13403 kB/run -  180.82 GB/s
ggml_metal_library_compile_pipeline: compiling pipeline: base = 'kernel_ssm_conv_f32_f32_4', name = 'kernel_ssm_conv_f32_f32_4'
ggml_metal_library_compile_pipeline: loaded kernel_ssm_conv_f32_f32_4                     0x12910b550 | th_max = 1024 | th_width =   32
  SSM_CONV(type=f32,ne_a=[4,3328,1,1],ne_b=[4,3328,1,1]):              98304 runs -    10.22 us/run -      117 kB/run -   10.91 GB/s
  Backend Metal: OK

Batch for both

ggml_metal_library_compile_pipeline: compiling pipeline: base = 'kernel_ssm_conv_f32_f32_b256', name = 'kernel_ssm_conv_f32_f32_b256'
ggml_metal_library_compile_pipeline: loaded kernel_ssm_conv_f32_f32_b256                  0x157b070e0 | th_max = 1024 | th_width =   32
  SSM_CONV(type=f32,ne_a=[515,3328,1,1],ne_b=[4,3328,1,1]):                    15024 runs -    70.58 us/run -    13403 kB/run -  181.11 GB/s
  SSM_CONV(type=f32,ne_a=[4,3328,1,1],ne_b=[4,3328,1,1]):              24576 runs -    59.17 us/run -      117 kB/run -    1.89 GB/s
  Backend Metal: OK

Non-Batch for both

ggml_metal_library_compile_pipeline: compiling pipeline: base = 'kernel_ssm_conv_f32_f32_4', name = 'kernel_ssm_conv_f32_f32_4'
ggml_metal_library_compile_pipeline: loaded kernel_ssm_conv_f32_f32_4                     0x135e075e0 | th_max = 1024 | th_width =   32
  SSM_CONV(type=f32,ne_a=[515,3328,1,1],ne_b=[4,3328,1,1]):                     2504 runs -  3801.51 us/run -    13403 kB/run -    3.36 GB/s
  SSM_CONV(type=f32,ne_a=[4,3328,1,1],ne_b=[4,3328,1,1]):              98304 runs -    10.26 us/run -      117 kB/run -   10.88 GB/s
  Backend Metal: OK

Given this, I think we should keep both versions dispatched on n_t like it is currently.

Branch: SSMKernelImprovements Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

gabe-l-hart · 2025-12-09T18:20:55Z

Some similar numbers for representative tests on SSM_CONV:

Without Optimizations

ggml_metal_library_compile_pipeline: compiling pipeline: base = 'kernel_ssm_scan_f32', name = 'kernel_ssm_scan_f32_nsg=4'
ggml_metal_library_compile_pipeline: loaded kernel_ssm_scan_f32_nsg=4                     0x149906a50 | th_max = 1024 | th_width =   32
  SSM_SCAN(type=f32,d_state=128,head_dim=64,n_head=48,n_group=1,n_seq_tokens=512,n_seqs=1):                     2102 runs -  2178.75 us/run -    15968 kB/run -    6.99 GB/s
  SSM_SCAN(type=f32,d_state=128,head_dim=64,n_head=48,n_group=1,n_seq_tokens=1,n_seqs=1):              40960 runs -    24.85 us/run -     3097 kB/run -  118.88 GB/s
  Backend Metal: OK

With Optimizations

ggml_metal_library_compile_pipeline: compiling pipeline: base = 'kernel_ssm_scan_f32', name = 'kernel_ssm_scan_f32_nsg=4'
ggml_metal_library_compile_pipeline: loaded kernel_ssm_scan_f32_nsg=4                     0x146408690 | th_max = 1024 | th_width =   32
  SSM_SCAN(type=f32,d_state=128,head_dim=64,n_head=48,n_group=1,n_seq_tokens=512,n_seqs=1):                     2102 runs -  1909.71 us/run -    15968 kB/run -    7.97 GB/s
  SSM_SCAN(type=f32,d_state=128,head_dim=64,n_head=48,n_group=1,n_seq_tokens=1,n_seqs=1):              40960 runs -    26.50 us/run -     3097 kB/run -  111.48 GB/s
  Backend Metal: OK

gabe-l-hart · 2025-12-09T18:22:53Z

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

+    x[0] = sumf;
+}
+
+// typedef decltype(kernel_ssm_conv_f32_f32_batched<1>) kernel_ssm_conv_batched_t;


Remove this

Branch: SSMKernelImprovements Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

gabe-l-hart · 2025-12-09T18:29:42Z

Another small speedup with a float4 version of the SSM_CONV batched impl:

ggml_metal_library_compile_pipeline: compiling pipeline: base = 'kernel_ssm_conv_f32_f32_batched_4', name = 'kernel_ssm_conv_f32_f32_batched_4_ssm_conv_bs=256'
ggml_metal_library_compile_pipeline: loaded kernel_ssm_conv_f32_f32_batched_4_ssm_conv_bs=256      0x126304280 | th_max = 1024 | th_width =   32
  SSM_CONV(type=f32,ne_a=[515,3328,1,1],ne_b=[4,3328,1,1]):                    15024 runs -    68.79 us/run -    13403 kB/run -  185.80 GB/s
ggml_metal_library_compile_pipeline: compiling pipeline: base = 'kernel_ssm_conv_f32_f32_4', name = 'kernel_ssm_conv_f32_f32_4'
ggml_metal_library_compile_pipeline: loaded kernel_ssm_conv_f32_f32_4                     0x12610d700 | th_max = 1024 | th_width =   32
  SSM_CONV(type=f32,ne_a=[4,3328,1,1],ne_b=[4,3328,1,1]):              98304 runs -    10.33 us/run -      117 kB/run -   10.80 GB/s
  Backend Metal: OK

gabe-l-hart · 2025-12-09T18:49:02Z

Hm, the failing test looks suspiciously related to this PR, but it's failing on CONV_2D which seems to pass just fine on my machine.

2025-12-09T18:44:28.7929820Z Failing tests:
2025-12-09T18:44:28.7930170Z   CONV_2D(ne_input=[1,1,25,2],ne_kernel=[2,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0)
2025-12-09T18:44:28.7930420Z   Backend Metal: �[1;31mFAIL�[0m

ggerganov

The failing tests seems like a fluke - should be safe to ignore

ggml/src/ggml-metal/ggml-metal-device.cpp

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

This brings in significant improvements to prefill performance for all models using the SSM_CONV and SSM_SCAN ops (granite4, jamba, falcon-h, nemotron-h, Qwen3 Next) on Apple Metal. See ggml-org/llama.cpp#17876 Branch: LlamaCPPMetalSSMImprovements Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

gabe-l-hart requested a review from ggerganov as a code owner December 8, 2025 23:52

loci-dev mentioned this pull request Dec 9, 2025

UPSTREAM PR #17876: metal: SSM kernel improvements auroralabs-loci/llama.cpp#494

Open

github-actions bot added ggml changes relating to the ggml tensor library for machine learning Apple Metal https://en.wikipedia.org/wiki/Metal_(API) labels Dec 9, 2025

ggerganov reviewed Dec 9, 2025

View reviewed changes

gabe-l-hart added 3 commits December 9, 2025 10:14

feat: Add a batched version of ssm_conv

da35bf1

This was done using Claude Code. It found a number of optimizations around how the threads were organized, resulting in a huge performance boost! Branch: Mamba2SSD Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

feat: Optimized SSM_SCAN kernel for metal

eab5301

This used Claude Code and resulted in a modest performance improvement while maintaining correctness. Branch: Mamba2SSD Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

test: Add test-backend-ops perf tests for SSM_CONV

c3090ec

Branch: SSMKernelImprovements Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

gabe-l-hart added 2 commits December 9, 2025 10:24

test: Real representitive tests for SSM_CONV

78a232c

Branch: SSMKernelImprovements Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

refactor: Use function constant for ssm_conv batch size

da044cd

Branch: SSMKernelImprovements Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

gabe-l-hart force-pushed the SSMKernelImprovements branch from 427ae08 to da044cd Compare December 9, 2025 18:04

test: backend op tests for ssm_scan from granite4 1b-h

d6c812d

Branch: SSMKernelImprovements Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

gabe-l-hart commented Dec 9, 2025

View reviewed changes

gabe-l-hart added 2 commits December 9, 2025 11:23

style: remove commented out templates

b2bd438

Branch: SSMKernelImprovements Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

feat: float4 version of ssm_conv_batched

07e3c93

Branch: SSMKernelImprovements Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

ggerganov approved these changes Dec 9, 2025

View reviewed changes

ggml/src/ggml-metal/ggml-metal-device.cpp Show resolved Hide resolved

fix: Add missing ggml_metal_cv_free

b7284e8

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

ggerganov merged commit 086a63e into ggml-org:master Dec 9, 2025
62 of 69 checks passed

gabe-l-hart deleted the SSMKernelImprovements branch December 9, 2025 19:30

github-actions bot added the testing Everything test related label Dec 9, 2025

gabe-l-hart mentioned this pull request Dec 10, 2025

feat: llama.cpp bump (17f7f4) for SSM performance improvements ollama/ollama#13408

Open

	constant short FC_mul_mv_nsg [[function_constant(FC_MUL_MV + 0)]];
	constant short FC_mul_mv_nxpsg [[function_constant(FC_MUL_MV + 1)]];

	template<typename block_q_type, short NR0, typename args_t>
	void mul_vec_q_n_f32_impl(
	args_t args,
	device const char * src0,
	device const char * src1,
	device char * dst,
	threadgroup char * shmem,
	uint3 tgpig,
	ushort tiisg,
	ushort sgitg) {
	const short NSG = FC_mul_mv_nsg;


	ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv_ext(ggml_metal_library_t lib, ggml_type tsrc0, ggml_type tsrc1, int nsg, int nxpsg, int r1ptg) {
	char base[256];
	char name[256];

	snprintf(base, 256, "kernel_mul_mv_ext_%s_%s_r1_%d", ggml_type_name(tsrc0), ggml_type_name(tsrc1), r1ptg);
	snprintf(name, 256, "%s_nsg=%d_nxpsg=%d", base, nsg, nxpsg);

	ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
	if (!res.pipeline) {
	ggml_metal_cv_t cv = ggml_metal_cv_init();

	ggml_metal_cv_set_int16(cv, nsg, FC_MUL_MV + 0);
	ggml_metal_cv_set_int16(cv, nxpsg, FC_MUL_MV + 1);

	res = ggml_metal_library_compile_pipeline(lib, base, name, cv);

	ggml_metal_cv_free(cv);
	}

	return res;
	}

metal: SSM kernel improvements #17876

metal: SSM kernel improvements #17876

Conversation

gabe-l-hart commented Dec 8, 2025

Description

Changes

Performance

Baseline (c8554b6)

SSM_CONV improvements

SSM_CONV + SSM_SCAN improvements

Uh oh!

jeffbolznv commented Dec 9, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

gabe-l-hart commented Dec 9, 2025

Test results for test-backend-ops perf -o SSM_CONV -b Metal

Baseline (non-batch)

Batched for n_t > 1

Batched always

Uh oh!

jeffbolznv commented Dec 9, 2025

Uh oh!

ggerganov commented Dec 9, 2025

Uh oh!

gabe-l-hart commented Dec 9, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

gabe-l-hart commented Dec 9, 2025

Batched

Un-Batched

Uh oh!

gabe-l-hart commented Dec 9, 2025

Batch for prefill only

Batch for both

Non-Batch for both

Uh oh!

gabe-l-hart commented Dec 9, 2025

Without Optimizations

With Optimizations

Uh oh!

Choose a reason for hiding this comment

Uh oh!

gabe-l-hart commented Dec 9, 2025

Uh oh!

gabe-l-hart commented Dec 9, 2025

Uh oh!

ggerganov left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Baseline (`c8554b6`)

Test results for `test-backend-ops perf -o SSM_CONV -b Metal`

gabe-l-hart commented Dec 9, 2025 •

edited

Loading