From: Tarek Dakhran <redacted>
Date: Fri, 11 Jul 2025 18:27:01 +0000 (+0200)
Subject: model : support LiquidAI LFM2 hybrid family (llama/14620)
X-Git-Tag: upstream/0.0.2309~5
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=5d6c37b4b23117bb54e1969963d97be4859d3f85;p=pkg%2Fggml%2Fsources%2Fggml

model : support LiquidAI LFM2 hybrid family (llama/14620)

**Important**
LFM2 was [merged ](https://github.com/huggingface/transformers/pull/39340)into transformers, but has not yet been released.
To convert into gguf, install transformers from source
```shell
pip install "transformers @ git+https://github.com/huggingface/transformers.git@main"
```
---

diff --git a/src/ggml-cuda/ssm-conv.cu b/src/ggml-cuda/ssm-conv.cu
index f6375719..41979733 100644
--- a/src/ggml-cuda/ssm-conv.cu
+++ b/src/ggml-cuda/ssm-conv.cu
@@ -107,8 +107,11 @@ static void ssm_conv_f32_cuda(const float * src0, const float * src1, const int
         if (nc == 4) {
             ssm_conv_f32<threads, 4><<<blocks, threads, 0, stream>>>(src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1,
                                                                      dst, dst_nb0, dst_nb1, dst_nb2, n_t);
+        } else if (nc == 3) {
+            ssm_conv_f32<threads, 3><<<blocks, threads, 0, stream>>>(src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1,
+                                                                     dst, dst_nb0, dst_nb1, dst_nb2, n_t);
         } else {
-            GGML_ABORT("Only support kernel size = 4  now.");
+            GGML_ABORT("Only support kernel size = 3 or size = 4 right now.");
         }
     } else {
         if (nc == 4) {
@@ -116,8 +119,13 @@ static void ssm_conv_f32_cuda(const float * src0, const float * src1, const int
             dim3          blocks(n_s, (nr + threads - 1) / threads, (n_t + split_n_t - 1) / split_n_t);
             ssm_conv_long_token_f32<threads, 4, split_n_t><<<blocks, threads, 0, stream>>>(
                 src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1, dst, dst_nb0, dst_nb1, dst_nb2, n_t);
+        } else if (nc == 3) {
+            const int64_t split_n_t = 32;
+            dim3          blocks(n_s, (nr + threads - 1) / threads, (n_t + split_n_t - 1) / split_n_t);
+            ssm_conv_long_token_f32<threads, 3, split_n_t><<<blocks, threads, 0, stream>>>(
+                src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1, dst, dst_nb0, dst_nb1, dst_nb2, n_t);
         } else {
-            GGML_ABORT("Only support kernel size = 4 right now.");
+            GGML_ABORT("Only support kernel size = 3 or size = 4 right now.");
         }
     }
 }