Ext: Fix CUDA type cast

kingbri1 · kingbri1 · commit 0a7733110e46 · 2025-05-12T12:36:40.000-04:00
The __half_as_ushort function isn't present in cuda &lt; 12.4

Signed-off-by: kingbri &lt;8082010+kingbri1@users.noreply.github.com&gt;
diff --git a/exllamav2/exllamav2_ext/cuda/q_mlp.cu b/exllamav2/exllamav2_ext/cuda/q_mlp.cu
@@ -416,7 +416,7 @@ void QMoEMLP::forward_
 
         for (int i = 0; i < num_experts; i++)
         {
-            uint16_t w = __half_as_ushort(h_logits[i]);
+            uint16_t w = *reinterpret_cast<const uint16_t*>(&h_logits[i]);
             if (!w) continue;
 
             gemm_half_q_half_cuda(stream, cublas_handle, temp_state, w1[i], temp_a, rows, intermediate_size, columns, true, temp_dq, true, temp_logits + i, num_experts, false);

Original file line number	Diff line number	Diff line change
`@@ -416,7 +416,7 @@ void QMoEMLP::forward_`
`416`	`416`
`417`	`417`	`for (int i = 0; i < num_experts; i++)`
`418`	`418`	`{`
`419`		`- uint16_t w = __half_as_ushort(h_logits[i]);`
	`419`	`+ uint16_t w = reinterpret_cast<const uint16_t>(&h_logits[i]);`
`420`	`420`	`if (!w) continue;`
`421`	`421`
`422`	`422`	`gemm_half_q_half_cuda(stream, cublas_handle, temp_state, w1[i], temp_a, rows, intermediate_size, columns, true, temp_dq, true, temp_logits + i, num_experts, false);`