|
|
@@ -1,14 +1,18 @@
|
|
|
# GLM-5.3-Flash on 2× DGX Spark — aidendle94 sparkrun image (TP=2).
|
|
|
#
|
|
|
# aiden's 4x guide (https://aidenle.com/recipes/glm-5-3-flash-4x-dgx-spark)
|
|
|
-# translated to our 2-node ds4 launcher: same image (production-1.1 — SM90
|
|
|
+# translated to our 2-node ds4 launcher: same image (production-1.2 — SM90
|
|
|
# NoPE-MLA-to-capability-12 route + three functional fixes + DFlash2 overlay),
|
|
|
# same serve flags, TP2 instead of TP4.
|
|
|
#
|
|
|
-# Image: aidendle94/sparkrun-vllm-glm53-gb10:production-1.1 (Docker Hub).
|
|
|
+# Image: aidendle94/sparkrun-vllm-glm53-gb10:production-1.2 (Docker Hub).
|
|
|
# production-1.0 = MTP-only fallback; 1.1 ships the DFlash2 drafter overlay
|
|
|
# (drafter package + aux-capture patch + KV-page LCM patch + backend-registry
|
|
|
-# patch). Grep the log for "B12X PCIe|DFlash" style markers per his guide.
|
|
|
+# patch); 1.2 adds a fused Triton mHC prefill kernel (post-mix + pre-norm
|
|
|
+# GEMM + sinkhorn pre-mix + RMSNorm in one pass; aiden: 1.37x on the mHC
|
|
|
+# section at 4K tokens, prefill ≥1024 tokens, hc_mult=4). Same wheel and
|
|
|
+# overlays otherwise. VLLM_MHC_FUSED_LARGE=0 reverts to the 1.1 path.
|
|
|
+# Grep the log for "B12X PCIe|DFlash" style markers per his guide.
|
|
|
#
|
|
|
# Weights: LibertAIDAI/GLM-5.3-Flash-NVFP4 (194 GB, 97 GiB/node at TP2; cached
|
|
|
# on both nodes). aiden: NVFP4 is ~23% faster than unsloth FP8 (which needs 4
|
|
|
@@ -32,7 +36,7 @@
|
|
|
|
|
|
engine = "vllm"
|
|
|
name = "GLM-5.3-Flash aiden image (vLLM, 2x Spark, DFlash2-7)"
|
|
|
-docker_image = "aidendle94/sparkrun-vllm-glm53-gb10:production-1.1"
|
|
|
+docker_image = "aidendle94/sparkrun-vllm-glm53-gb10:production-1.2"
|
|
|
model = "LibertAIDAI/GLM-5.3-Flash-NVFP4"
|
|
|
required_nodes = 2
|
|
|
tensor_parallel_size = 2
|
