turn off plugin on deepseek models to prevent error

philipkiely-baseten · philipkiely-baseten · commit 743ae2103b9f · 2025-03-06T15:40:38.000-08:00
diff --git a/deepseek/engine-deepseek-r1-distill-qwen-14b/config.yaml b/deepseek/engine-deepseek-r1-distill-qwen-14b/config.yaml
@@ -38,7 +38,7 @@ trt_llm:
     tensor_parallel_count: 1
     plugin_configuration:
       use_paged_context_fmha: true
-      use_fp8_context_fmha: true
+      use_fp8_context_fmha: false
       paged_kv_cache: true
   runtime:
     batch_scheduler_policy: max_utilization
diff --git a/deepseek/engine-deepseek-r1-distill-qwen-32b/config.yaml b/deepseek/engine-deepseek-r1-distill-qwen-32b/config.yaml
@@ -38,7 +38,7 @@ trt_llm:
     tensor_parallel_count: 1
     plugin_configuration:
       use_paged_context_fmha: true
-      use_fp8_context_fmha: true
+      use_fp8_context_fmha: false
       paged_kv_cache: true
   runtime:
     batch_scheduler_policy: max_utilization
