glm53 aiden: local image (prod-1.1 + instanttensor), MTP spec, 512K ctx

- new docker-images/glm53-flash-aiden: aiden prod-1.1 + instanttensor 0.1.9
- aiden-4x: local image, zai-org model, MTP k=5 (dflash2 commented out),
  --load-format instanttensor, reasoning_effort max
- aiden-4x-labdrafter: local image, lab NVFP4 repo, MTP k=3, 512K ctx,
  flashinfer_cutlass MoE (block-size 2304 commented), no eager, MAX_JOBS=2
- dflash2-mxfp8: 512K ctx, --load-format instanttensor
- r12: BUILD_JOBS=4 pin
- drop data/benchmarks.sqlite

docker-images/glm53-flash-aiden/Dockerfile

@@ -0,0 +1,3 @@
+FROM aidendle94/sparkrun-vllm-glm53-gb10:production-1.1
+
+RUN pip install --no-cache-dir --no-deps --force-reinstall instanttensor==0.1.9

docker-images/glm53-flash-aiden/build.sh

@@ -0,0 +1,7 @@
+#!/usr/bin/env bash
+set -euo pipefail
+cd "$(dirname "$0")"
+TAG="${TAG:-sparkrun-vllm-glm53-flash-aiden:glm53}"
+echo "=== Building ${TAG} ==="
+docker build --progress=plain -t "${TAG}" . 2>&1
+echo "=== Done: ${TAG} ==="

docker-images/glm53-jovian-r12/versions.env

@@ -42,3 +42,4 @@ DEEPGEMM_REF=8b1392b978f5a03c828dd1711090d7fb50958b8a
 
 # Target GPU architecture (DGX Spark = sm_121a)
 TORCH_CUDA_ARCH_LIST=12.1a
+BUILD_JOBS=4

recipes/glm53-flash-aiden-4x-labdrafter/recipe.toml

@@ -37,8 +37,9 @@
 
 engine = "vllm"
 name = "GLM-5.3-Flash aiden image (vLLM, 4x Spark, LAB DFlash2-7, NVFP4)"
-docker_image = "aidendle94/sparkrun-vllm-glm53-gb10:production-1.1"
-model = "LibertAIDAI/GLM-5.3-Flash-NVFP4"
+docker_image = "sparkrun-vllm-glm53-flash-aiden:glm53"
+docker_image_dir = "glm53-flash-aiden"
+model = "local-inference-lab/GLM-5.3-Flash-NVFP4"
 required_nodes = 4
 tensor_parallel_size = 4
 pipeline_parallel_size = 1
@@ -57,18 +58,21 @@ args = [
     "--default-chat-template-kwargs", """
 '{"enable_thinking": true, "reasoning_effort": "high"}'""",
     "--kv-cache-dtype", "fp8_e4m3",
-    "--block-size", "2304",
-    "--moe-backend", "marlin",
-    "--max-model-len", "393216",
+#    "--block-size", "2304",
+    "--moe-backend", "flashinfer_cutlass",
+    "--max-model-len", "524288",
     "--max-num-batched-tokens", "8192",
     "--max-num-seqs", "16",
     "--enable-chunked-prefill",
-    "--enforce-eager",
+#    "--enforce-eager",
     "--gpu-memory-utilization", "0.85",
     "--kv-cache-memory-bytes", "17179869184",
     "--speculative-config", """
-'{"method":"dflash","model":"local-inference-lab/GLM-5.3-Flash-DFlash2-bf16","num_speculative_tokens":7}'""",
+'{"method":"mtp","num_speculative_tokens":3}'""",
+#    "--speculative-config", """
+#'{"method":"dflash","model":"local-inference-lab/GLM-5.3-Flash-DFlash2-bf16","num_speculative_tokens":7}'""",
     "--language-model-only",
+    "--load-format", "instanttensor",
     # image requests need the mm template shipped/in-wallpaper on aiden's rigs:
     # "--chat-template", "<path-to>/chat_template_mm.jinja",
 ]
@@ -85,3 +89,4 @@ VLLM_ENGINE_READY_TIMEOUT_S = "3600"
 VLLM_ALLOW_LONG_MAX_MODEL_LEN = "1"
 VLLM_NO_USAGE_STATS = "1"
 OMP_NUM_THREADS = "2"
+MAX_JOBS = "2"

recipes/glm53-flash-aiden-4x/recipe.toml

@@ -24,8 +24,9 @@
 
 engine = "vllm"
 name = "GLM-5.3-Flash aiden image (vLLM, 4x Spark, DFlash2-7, FP8) [DEFERRED]"
-docker_image = "aidendle94/sparkrun-vllm-glm53-gb10:production-1.1"
-model = "LibertAIDAI/GLM-5.3-Flash-NVFP4"
+docker_image = "sparkrun-vllm-glm53-flash-aiden:glm53"
+docker_image_dir = "glm53-flash-aiden"
+model = "zai-org/GLM-5.3-Flash"
 required_nodes = 4
 tensor_parallel_size = 4
 pipeline_parallel_size = 1
@@ -42,7 +43,7 @@ args = [
     "--tool-call-parser", "glm47",
     "--enable-auto-tool-choice",
     "--default-chat-template-kwargs", """
-'{"enable_thinking": true, "reasoning_effort": "high"}'""",
+'{"enable_thinking": true, "reasoning_effort": "max"}'""",
     "--kv-cache-dtype", "fp8_e4m3",
     "--block-size", "2304",
     "--moe-backend", "marlin",
@@ -53,9 +54,12 @@ args = [
     "--enforce-eager",
     "--gpu-memory-utilization", "0.85",
     "--kv-cache-memory-bytes", "17179869184",
+#    "--speculative-config", """
+#'{"method":"dflash","model":"incoai/GLM-5.3-Flash-DFlash2","num_speculative_tokens":7}'""",
     "--speculative-config", """
-'{"method":"dflash","model":"incoai/GLM-5.3-Flash-DFlash2","num_speculative_tokens":7}'""",
+'{"method":"mtp","num_speculative_tokens":5}'""",
     "--language-model-only",
+    "--load-format", "instanttensor",
     # image requests need the mm template shipped/in-wallpaper on aiden's rigs:
     # "--chat-template", "<path-to>/chat_template_mm.jinja",
 ]

recipes/glm53-flash-dflash2-mxfp8/recipe.toml

@@ -99,7 +99,7 @@ args = [
     "--kv-cache-memory-bytes", "10737418240",
     "--kv-cache-dtype", "fp8_ds_mla",
     "--block-size", "256",
-    "--max-model-len", "262144",
+    "--max-model-len", "524288",
     "--max-num-seqs", "8",
     "--max-num-batched-tokens", "4096",
     "--enable-chunked-prefill",
@@ -112,6 +112,7 @@ args = [
     "--speculative-config", """
 '{"method":"dflash","model":"local-inference-lab/GLM-5.3-Flash-DFlash2-MXFP8","num_speculative_tokens":7,"kv_cache_dtype":"auto"}'""",
     "--language-model-only",
+    "--load-format", "instanttensor",
     # See header: eager is the safe control on the day-0 base (graphs+spec
     # quality regression). The card's newer build runs FULL graphs.
     "--enforce-eager",