|
|
@@ -27,7 +27,7 @@
|
|
|
|
|
|
engine = "vllm"
|
|
|
name = "GLM-5.3-Flash aiden image (vLLM, 4x Spark, MTP-5, FP8)"
|
|
|
-docker_image = "aidendle94/sparkrun-vllm-glm53-gb10:production-1.2" # hub tag (parity with the benched 2x lane)
|
|
|
+docker_image = "sparkrun-vllm-glm53-flash-aiden:glm53" # production-1.2 + instanttensor (docker-images/glm53-flash-aiden/)
|
|
|
docker_image_dir = "glm53-flash-aiden"
|
|
|
model = "zai-org/GLM-5.3-Flash"
|
|
|
required_nodes = 4
|
|
|
@@ -61,8 +61,8 @@ args = [
|
|
|
#'{"method":"dflash","model":"incoai/GLM-5.3-Flash-DFlash2","num_speculative_tokens":7}'""",
|
|
|
"--speculative-config", """
|
|
|
'{"method":"mtp","num_speculative_tokens":5}'""",
|
|
|
- # No --load-format instanttensor: the production-1.2 image does not ship
|
|
|
- # the instanttensor package (the 2x lane also loads via plain safetensors).
|
|
|
+ # instanttensor comes from our extension layer (base image lacks it)
|
|
|
+ "--load-format", "instanttensor",
|
|
|
"--language-model-only",
|
|
|
# image requests need the mm template shipped/in-wallpaper on aiden's rigs:
|
|
|
# "--chat-template", "<path-to>/chat_template_mm.jinja",
|
