Bump GLM recipes to aidendle94 production-hybrid-1.6

Newer tag of the production-hybrid line (2026-07-26); image env identical
to 1.0 (MODEL_PATH, SPEC_TOKENS=5, VLLM_USE_V2_MODEL_RUNNER), so a
drop-in bump for the three recipes still pinned at 1.0. NF3 caveat kept:
1.6 still lacks the hybrid overlay files, revert to vllm-nf3-hybrid:probe
if it fails to boot.

recipes/glm52-int4-int8/recipe.toml

@@ -1,6 +1,6 @@
 # GLM-5.2 QuantTrio Int4-Int8Mix — 4× DGX Spark / GB10 (sm_121), TP=4.
 #
-# EXPERIMENT: Using aidendle94/sparkrun-vllm-ds4-gb10:production-hybrid-1.0
+# EXPERIMENT: Using aidendle94/sparkrun-vllm-ds4-gb10:production-hybrid-1.6
 # instead of locally-built sparkrun-vllm-ds4-gb10:glm52.
 # Pre-built on Docker Hub with CUDA 13.2.1, custom NCCL, b12x suite.
 #
@@ -18,7 +18,7 @@
 
 engine = "vllm"
 name = "GLM-5.2 Int4-Int8Mix (vLLM, DS4 image)"
-docker_image = "aidendle94/sparkrun-vllm-ds4-gb10:production-hybrid-1.0"
+docker_image = "aidendle94/sparkrun-vllm-ds4-gb10:production-hybrid-1.6"
 model = "QuantTrio/GLM-5.2-Int4-Int8Mix"
 required_nodes = 4
 tensor_parallel_size = 4

recipes/glm52-nf3-hybrid-context/recipe.toml

@@ -8,9 +8,9 @@
 #   327GB — top-64 experts/layer in NVFP4, remaining 192 in NF3 (3-bit),
 #   non-expert layers in MXFP8.
 #
-# EXPERIMENT: Using aidendle94/sparkrun-vllm-ds4-gb10:production-hybrid-1.0
+# EXPERIMENT: Using aidendle94/sparkrun-vllm-ds4-gb10:production-hybrid-1.6
 # instead of locally-built vllm-nf3-hybrid:probe. Pre-built on Docker Hub.
-# NOTE: production-hybrid-1.0 lacks NF3 hybrid overlay files
+# NOTE: production-hybrid-1.6 lacks NF3 hybrid overlay files
 # (hybrid_loader.py, nf3_kernel.py, nvfp4_kernel.py, mxfp8_tier.json).
 # If it fails, revert to vllm-nf3-hybrid:probe.
 #
@@ -40,7 +40,7 @@
 
 engine = "vllm"
 name = "GLM-5.2 NF3-Hybrid CONTEXT LANE (vLLM)"
-docker_image = "aidendle94/sparkrun-vllm-ds4-gb10:production-hybrid-1.0"
+docker_image = "aidendle94/sparkrun-vllm-ds4-gb10:production-hybrid-1.6"
 model = "madeby561/GLM-5.2-MXFP8-NVFP4-NF3-Hybrid"
 required_nodes = 4
 tensor_parallel_size = 4

recipes/glm52-nf3-hybrid-fast/recipe.toml

@@ -8,9 +8,9 @@
 #   327GB — top-64 experts/layer in NVFP4, remaining 192 in NF3 (3-bit),
 #   non-expert layers in MXFP8.
 #
-# EXPERIMENT: Using aidendle94/sparkrun-vllm-ds4-gb10:production-hybrid-1.0
+# EXPERIMENT: Using aidendle94/sparkrun-vllm-ds4-gb10:production-hybrid-1.6
 # instead of locally-built vllm-nf3-hybrid:probe. Pre-built on Docker Hub.
-# NOTE: production-hybrid-1.0 lacks NF3 hybrid overlay files
+# NOTE: production-hybrid-1.6 lacks NF3 hybrid overlay files
 # (hybrid_loader.py, nf3_kernel.py, nvfp4_kernel.py, mxfp8_tier.json).
 # If it fails, revert to vllm-nf3-hybrid:probe.
 #
@@ -40,7 +40,7 @@
 
 engine = "vllm"
 name = "GLM-5.2 NF3-Hybrid FAST LANE (vLLM)"
-docker_image = "aidendle94/sparkrun-vllm-ds4-gb10:production-hybrid-1.0"
+docker_image = "aidendle94/sparkrun-vllm-ds4-gb10:production-hybrid-1.6"
 model = "madeby561/GLM-5.2-MXFP8-NVFP4-NF3-Hybrid"
 required_nodes = 4
 tensor_parallel_size = 4