Engine sparkglm/atlas-20261009: copy drafts on, streamed GLM text keeps tool/user/assistant, opt-in KV shard

.env.example

@@ -9,8 +9,10 @@ WORKER=you@192.0.2.2
 # path is used on both Sparks. Default: $HOME/models/sparkglm
 #MODEL_ROOT=/srv/models/sparkglm
 
-# 4x512k (default): 4 requests x 512K context. 8x128k: 8 requests x 128K.
-#PROFILE=4x512k
+# 4x1m (default): 4 requests x 1M context, the model's full window; the
+# requests share one KV pool (~2.62M tokens with the shard and the carveout at
+# 0.91; a 925K-token needle prompt is found). 4x512k: 4 x 512K. 8x128k: 8 x 128K.
+#PROFILE=4x1m
 
 # Share of each Spark's unified memory for the engine (0.80-0.95, default 0.88).
 # On Sparks that run nothing else, 0.91 roughly doubles the KV pool that holds
@@ -32,6 +34,13 @@ WORKER=you@192.0.2.2
 # /run/lock/sparkglm. Default off; set 0 or delete the line to roll back.
 #DISPLAY_CARVEOUT=1
 
+# The KV cache is split between the two Sparks: each stores the latents of half
+# the blocks, so the shared pool holds about 1.8x the tokens. With the display
+# carveout at 0.91: ~2.62M tokens (1.36M unsplit), lowest free memory 4.5 GB in
+# a four-session 204K fill. On by default; the disk prefix cache cannot run
+# beside it yet, so setting PREFIX_CACHE_DIR turns it off. 0 turns it off.
+#KV_SHARD=0
+
 # Build the image from source instead of pulling ghcr.io/enntity/atlas-sparkglm.
 #BUILD=1
 # Parallel compile jobs for a source build; lower it if the build runs out of memory.

README.md

@@ -229,7 +229,7 @@ comparison videos.
 ## The engine
 
 Atlas here is [`Enntity/atlas`](https://github.com/Enntity/atlas) branch
-`sparkglm/atlas-20261007-strings` (pinned in
+`sparkglm/atlas-20261009` (pinned in
 [`install/atlas-source.json`](install/atlas-source.json)), built from three
 layers:
 

docs/INSTALL-MANUAL.md

@@ -5,7 +5,7 @@ time. Run steps 1–4 on **both** Sparks with the same `MODEL_ROOT`.
 
 | Component | Pin |
 |---|---|
-| Engine | [`Enntity/atlas`](https://github.com/Enntity/atlas) `sparkglm/atlas-20261007-strings` @ `7838b34c` ([`install/atlas-source.json`](../install/atlas-source.json)) |
+| Engine | [`Enntity/atlas`](https://github.com/Enntity/atlas) `sparkglm/atlas-20261009` @ `e034500d` ([`install/atlas-source.json`](../install/atlas-source.json)) |
 | Model | [`nvidia/GLM-5.3-Flash-NVFP4`](https://huggingface.co/nvidia/GLM-5.3-Flash-NVFP4) @ `423acf37583782c51c142d145aef733d72943d93` (MIT) |
 | Drafter | [`incoai/GLM-5.3-Flash-DFlash2`](https://huggingface.co/incoai/GLM-5.3-Flash-DFlash2) @ `7d74cdd881ed7e32c31175984a67823127b66cfe` (CC BY-NC-ND 4.0) |
 | Native dependencies | FlashInfer `8eccd0c1`, CUTLASS `cf064d2e`, FlashKDA ([`install/flash_kda/`](../install/flash_kda/)), Rust 1.93.1, CUDA 13.0 |

install/atlas-source.json

@@ -2,24 +2,24 @@
   "schema": "sparkglm.atlas-engine-source/v2",
   "note": "The Atlas engine is built from the public Enntity/atlas fork at this exact commit; build.sh refuses any other tree.",
   "repository": "https://github.com/Enntity/atlas.git",
-  "branch": "sparkglm/atlas-20261007-strings",
-  "commit": "7838b34c18785ea876a758eb4806d005d3fcda38",
-  "tree": "d20749a55db35bae54a75ab2d67c5ce3def2021f",
+  "branch": "sparkglm/atlas-20261009",
+  "commit": "e034500d858742d131529e56179453e8afb45dce",
+  "tree": "edc66e7c3ca234e6aac3026c2b8242119dc42183",
   "layers": {
     "atlas_inf_main_base": "6a3d24ecb891adac1fac8edeef92de29bb6f3e15",
     "upstream_candidate": {
-      "branch": "upstream/glm53-flash-20261007-strings",
-      "commit": "89f2927b4b729db8da3925449d3e252d321d13db",
-      "note": "upstream/glm53-flash-20261004-gram (619da6ff: strict structured output, linear-cost grammar compile, masked DFlash verify for strict grammars) plus two tool-use fixes: the GLM (poolside_v1) tool parser keeps argument text and types it from the tool schema, so string parameters are no longer turned into booleans, numbers or lists; and the GLM template omits the reasoning-effort header on the empty think block (thinking off with tools)."
+      "branch": "upstream/glm53-flash-20261009",
+      "commit": "576ad34e",
+      "note": "upstream/glm53-flash-20261007-strings (GLM tool arguments keep their schema types; no effort header on the empty think block) plus copy (prompt-lookup) drafts beside DFlash2 (ATLAS_DFLASH_COPY_DRAFTS), the bare role-literal strip limited to ChatML models (GLM streamed replies lost the words tool/user/assistant), and Poolside tool-call headers held until the call validates."
     },
     "sparkglm_only": {
       "commit": "72588541a466f53cb9d38afd74c515e14ab601db",
-      "note": "The unchanged FlashKDA/native sparse-MLA prefill bridges and GB10 display-carveout KV placement from fdd2f965, then the carveout takes latent KV pools only (ATLAS_KV_CARVEOUT_ORDER, default latent): the GPU maps the carveout L2-uncached, and sparse-index buffers there slowed 64K cold prefill by 7% (Enntity/sparkglm#30). Fork-only; used only with DISPLAY_CARVEOUT=1."
+      "note": "The unchanged FlashKDA/native sparse-MLA prefill bridges and GB10 display-carveout KV placement from fdd2f965, then the carveout takes latent KV pools only (ATLAS_KV_CARVEOUT_ORDER, default latent): the GPU maps the carveout L2-uncached, and sparse-index buffers there slowed 64K cold prefill by 7% (Enntity/sparkglm#30). Fork-only; used only with DISPLAY_CARVEOUT=1. Plus: sharded latent pools (ATLAS_GLM_KV_SHARD, opt-in through SPARKGLM_KV_SHARD) size and place in the display carveout."
     },
     "integration": {
-      "branch": "fix/poolside-string-args",
-      "commit": "72588541 + 1a39c032 + 7838b34c",
-      "note": "This pin is sparkglm/atlas-20261005-cvplace (72588541) plus the two upstream-candidate fixes, cherry-picked unchanged; on the MLPerf edge-agentic BFCL leg the parser fix moves 83.72 to 88.74 overall (Enntity/sparkglm-research experiments/2026-10-06-mlperf-edge-agentic)."
+      "branch": "sparkglm/atlas-20261009",
+      "commit": "7838b34c + 9 cherry-picks",
+      "note": "sparkglm/atlas-20261007-strings plus the kv-shard carveout placement (exp/kvshard-20261008), copy drafts (exp/copy-drafts-20261008), the role-literal fix (fix/role-literal-strip-chatml) and the Poolside header hold (sparkqwen/rc15-fixes 25d14fcb), cherry-picked with -x."
     }
   },
   "kernel_target": "glm-5.3-flash",

install/profiles/4x1m.json

@@ -0,0 +1,223 @@
+{
+ "schema": "sparkglm.atlas-install-profile/v1",
+ "status": "candidate_2026-10-09",
+ "environment": {
+  "ATLAS_GLM_K3_MLA_QUERY_BATCHM": "1",
+  "ATLAS_GLM_K3_MLA_QUERY_COMPARE": "0",
+  "ATLAS_GLM_SPARSE_DECODE_SPLIT": "1",
+  "ATLAS_GLM_K3_MLA_O_BATCHM": "1",
+  "ATLAS_GLM_K3_HEAD_BATCHM": "0",
+  "ATLAS_GLM_C3_GROUPED_MOE": "1",
+  "ATLAS_GLM_SPARSE_DECODE_TC": "1",
+  "ATLAS_GLM_SPARSE_NATIVE": "1",
+  "ATLAS_GLM_SPARSE_NATIVE_LIBRARY": "/opt/atlas/native/libatlas_glm_sparse_native.so",
+  "ATLAS_GLM_HC_TF32_PREWARM": "1",
+  "ATLAS_KDA_FLASH_PREFILL": "1",
+  "ATLAS_KDA_FLASH_LIBRARY": "/opt/atlas/native/libatlas_mango_flash.so",
+  "ATLAS_EP_PROTOCOL": "v2",
+  "ATLAS_GLM_MTP_LONG_CONTEXT": "1",
+  "ATLAS_CUBLAS_GEMM": "1",
+  "ATLAS_DEBUG_NO_GRAPH": "0",
+  "ATLAS_EP_GRAPHS": "1",
+  "ATLAS_GLM_C4_DECODE": "0",
+  "ATLAS_GLM_C4_SPARSE": "0",
+  "ATLAS_GLM_INDEPENDENT_DECODE": "0",
+  "ATLAS_GLM_INDEX_ROW_GROUP": "8",
+  "ATLAS_GLM_INDEX_WMMA": "1",
+  "ATLAS_GLM_K5_BATCHED_CONV_SNAPSHOT": "1",
+  "ATLAS_GLM_K5_BATCHED_RECURRENT_SNAPSHOT": "1",
+  "ATLAS_GLM_K5_BATCHED_SHARED": "0",
+  "ATLAS_GLM_K5_COMPACT_MOE": "1",
+  "ATLAS_GLM_K5_DENSE_EXACT": "1",
+  "ATLAS_GLM_K5_FUSED_COMPACT_GATE_UP": "1",
+  "ATLAS_GLM_K5_FUSED_DENSE_PAIRS": "0",
+  "ATLAS_GLM_K5_FUSED_DENSE_TRIPLE": "1",
+  "ATLAS_GLM_K5_FUSED_MOE_HC": "1",
+  "ATLAS_GLM_K5_FUSED_QKV": "1",
+  "ATLAS_GLM_K5_FUSED_SHARED_GATE_UP": "1",
+  "ATLAS_GLM_K5_FUSED_TP_HC": "1",
+  "ATLAS_GLM_K5_GROUPED_MOE": "1",
+  "ATLAS_GLM_K5_HC_CUBLAS": "1",
+  "ATLAS_GLM_K5_ROUTER_M5": "1",
+  "ATLAS_GLM_M5_ROUTER_BN4": "1",
+  "ATLAS_GLM_M5_ROUTER_BN4_VERIFY": "0",
+  "ATLAS_GLM_MOE_GATE_UP_M16": "1",
+  "ATLAS_GLM_MOE_GATE_UP_M16_VERIFY": "0",
+  "ATLAS_GLM_MTP_ALL_GATHER": "0",
+  "ATLAS_GLM_MTP_BATCHED_PREFILL": "0",
+  "ATLAS_GLM_MTP_BF16_DRAFTS": "0",
+  "ATLAS_GLM_MTP_DISTRIBUTED": "0",
+  "ATLAS_GLM_MTP_DISTRIBUTED_ARGMAX": "0",
+  "ATLAS_GLM_MTP_FUSED_EH_CHECK": "0",
+  "ATLAS_GLM_MTP_FUSED_EH_NORM": "0",
+  "ATLAS_GLM_MTP_HIDDEN_TRACE": "0",
+  "ATLAS_GLM_MTP_K5_LEDGER": "0",
+  "ATLAS_GLM_MTP_KV_REPAIR_VERIFY": "0",
+  "ATLAS_GLM_MTP_NVFP4_EH": "0",
+  "ATLAS_GLM_MTP_NVFP4_WO": "0",
+  "ATLAS_GLM_MTP_PROFILE": "0",
+  "ATLAS_GLM_MTP_REPAIR": "0",
+  "ATLAS_GLM_MTP_SERIAL_PREFILL": "0",
+  "ATLAS_GLM_MULTI_SEQ_SPARSE": "0",
+  "ATLAS_GLM_MULTI_SEQ_SPARSE_GRAPHS": "0",
+  "ATLAS_GLM_PAGED_PREFILL_BF16_GEMM": "1",
+  "ATLAS_GLM_SPARSE_HEAD_GROUP": "8",
+  "ATLAS_GLM_SSM_NORMALIZE": "0",
+  "ATLAS_GLM_TARGET_SHARED_FP8": "1",
+  "ATLAS_GLM_TARGET_SHARED_FP8_VERIFY": "0",
+  "ATLAS_GLM_MLA_MULTI_SEQ": "1",
+  "ATLAS_GLM_MLA_BATCH23": "1",
+  "ATLAS_GLM_MTP1_VERIFY_GRAPH": "0",
+  "ATLAS_GLM_TP_VERIFY_GRAPH": "0",
+  "ATLAS_HC_CUBLAS_PREFILL": "1",
+  "ATLAS_KDA_REGRESIDENT_PREFILL": "1",
+  "ATLAS_GLM_KV_CAP_TO_CONTEXTS": "0",
+  "ATLAS_MOE_GROUPED_CUTLASS": "0",
+  "ATLAS_MOE_PREFILL_EXACT_TILES": "1",
+  "ATLAS_MOE_PREFILL_MAX_LOAD_FACTOR": "8",
+  "ATLAS_MOE_SHARED_REDUCE_OVERLAP": "1",
+  "ATLAS_MTP_DRAFT_CONF": "0",
+  "ATLAS_MTP_SINGLE_DEPTH_ADAPT": "0",
+  "ATLAS_MTP_SPEC_THINK": "0",
+  "ATLAS_NO_DECODE_GRAPHS": "0",
+  "ATLAS_NO_DECODE_GRAPHS_MULTISEQ": "0",
+  "ATLAS_NVFP4_FUSED_SILU_QUANT": "1",
+  "ATLAS_NVFP4_MMQ_MOE": "0",
+  "ATLAS_NVFP4_PREQUANT_MOE": "1",
+  "ATLAS_NVFP4_VECSCALE": "1",
+  "ATLAS_UNIFIED_MOE_LAYOUT": "1",
+  "CUDA_CACHE_PATH": "/atlas-cuda-cache",
+  "LD_LIBRARY_PATH": "/opt/atlas/native:/usr/local/nvidia/lib:/usr/local/nvidia/lib64:/usr/local/cuda/lib64",
+  "NCCL_ALGO": "Ring",
+  "NCCL_BUFFSIZE": "33554432",
+  "NCCL_CUMEM_ENABLE": "0",
+  "NCCL_CUMEM_HOST_ENABLE": "0",
+  "NCCL_DEBUG": "WARN",
+  "NCCL_DMABUF_ENABLE": "0",
+  "NCCL_IB_ADDR_FAMILY": "AF_INET",
+  "NCCL_IB_DISABLE": "0",
+  "NCCL_CROSS_NIC": "0",
+  "NCCL_IB_RETRY_CNT": "7",
+  "NCCL_IB_ROCE_VERSION_NUM": "2",
+  "NCCL_IB_TIMEOUT": "22",
+  "NCCL_MAX_NCHANNELS": "2",
+  "NCCL_MIN_NCHANNELS": "1",
+  "NCCL_NET": "IB",
+  "NCCL_NET_GDR_C2C": "0",
+  "NCCL_NET_GDR_LEVEL": "0",
+  "NCCL_NVLS_ENABLE": "0",
+  "NCCL_PROTO": "Simple",
+  "PATH": "/usr/local/nvidia/bin:/usr/local/cuda/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin",
+  "RUST_LOG": "info",
+  "ATLAS_GLM_PAGED_PREFILL_MLA_GEMM": "1",
+  "ATLAS_GLM_SPARSE_PREFILL_TC": "1",
+  "ATLAS_GLM_PREFILL_2048": "0",
+  "ATLAS_GLM_DENSE_PREFILL_BF16": "2",
+  "ATLAS_GLM_SPARSE_PREFILL_KV_REUSE": "1",
+  "ATLAS_GLM_PREFILL_4096": "1",
+  "ATLAS_GLM_ROUTER_PREFILL_BN32": "1",
+  "ATLAS_GLM_HC_PREFILL_VEC": "1",
+  "ATLAS_GLM_KDA_PREFILL_LT_FP8": "1",
+  "ATLAS_HC_GENERIC_MOE_ARM": "1",
+  "ATLAS_GLM_LONG_BATCH_VERIFY": "1",
+  "ATLAS_GLM_MTP_REPAIR_CONTEXT": "65536",
+  "ATLAS_GLM_HC_FUSED_PREFILL": "1",
+  "ATLAS_GLM_VERIFY_VOCAB_SPLIT": "1",
+  "ATLAS_GLM_DFLASH": "1",
+  "ATLAS_DFLASH_OPTION_B": "1",
+  "ATLAS_W4A16_TC": "1",
+  "ATLAS_GLM_ROUTER_PREFILL_CUBLAS": "1",
+  "ATLAS_GLM_HC_BF16": "1",
+  "ATLAS_DFLASH_MXFP8": "1",
+  "ATLAS_BF16_GEMM_CUTLASS": "1",
+  "ATLAS_GLM_HC_POST_MIX": "1",
+  "ATLAS_DFLASH_MULTI_DRAFT_CAP": "7",
+  "ATLAS_RDMA_ALLREDUCE": "1",
+  "ATLAS_GLM_PREFILL_SP": "1",
+  "ATLAS_GLM_LM_HEAD_MXFP8": "1",
+  "ATLAS_GLM_MLA_MXFP8": "1",
+  "ATLAS_DFLASH_MXFP8_HEAD": "1",
+  "ATLAS_DFLASH_NVFP4_HEAD": "1",
+  "ATLAS_MOE_PREQUANT_K128": "1",
+  "ATLAS_GLM_HC_DECODE_SEAM": "1",
+  "ATLAS_DFLASH_NVFP4_TC": "1",
+  "ATLAS_DFLASH_DROP_BF16": "1",
+  "ATLAS_DFLASH_ADAPTIVE_WIDTH": "1",
+  "ATLAS_DFLASH_SPEC_THINK": "1",
+  "ATLAS_GLM_SHARED_TP_SPLIT": "1",
+  "ATLAS_GLM_EXPERT_TP": "1",
+  "ATLAS_MARCONI_PREFILL_ONLY": "1",
+  "ATLAS_PREFIX_SUBBLOCK": "0",
+  "ATLAS_DFLASH_CTX_NVFP4": "1",
+  "ATLAS_PDL": "1",
+  "ATLAS_GLM_SPARSE_PREFILL_PIPE": "1",
+  "ATLAS_GLM_MOE_PREFILL_PERSIST": "1",
+  "ATLAS_GLM_MOE_UNPERMUTE_VEC": "1",
+  "ATLAS_GLM_INDEX_LOGITS_V2": "1",
+  "ATLAS_GLM_INDEX_SPLIT": "1",
+  "ATLAS_GLM_ROUTER_PREFILL_CUTLASS": "1",
+  "ATLAS_GLM_MOE_DECODE_M16": "1",
+  "ATLAS_GLM_MOE_DOWN_ZSKIP": "1",
+  "ATLAS_GLM_PC_EVICT": "1",
+  "ATLAS_GLM_PC_BRANCH": "1",
+  "ATLAS_GLM_WARM_SKIP_CACHED": "1",
+  "ATLAS_GLM_WARM_CHUNK_RUN": "1",
+  "ATLAS_PREFILL_SRPT": "1",
+  "ATLAS_DFLASH_FIRST_APPEND": "none",
+  "ATLAS_GLM_CMD_RDMA": "1",
+  "ATLAS_GLM_MOE_DECODE_STREAM": "1",
+  "ATLAS_GLM_DRAFT_TP": "1",
+  "ATLAS_GLM_DECODE_FUSE": "1",
+  "ATLAS_GLM_DECODE_GEMV_BATCH": "1",
+  "ATLAS_GLM_MOE_DECODE_L2PF": "1",
+  "ATLAS_GLM_MOE_STREAM_NOSYNC": "1",
+  "ATLAS_GLM_DRAFT_TP_BATCH": "1",
+  "ATLAS_RDMA_ONESHOT": "1",
+  "ATLAS_RDMA_PAIR_CHAIN": "1",
+  "ATLAS_GLM_HC_SEAM_ILP": "3",
+  "ATLAS_GLM_DRAFT_TP_CTX": "1",
+  "ATLAS_DFLASH_CTX_ASYNC_POS": "1",
+  "ATLAS_GLM_STRICT_SPEC": "1",
+  "ATLAS_DFLASH_CONF_WIDTH": "1",
+  "ATLAS_USE_HF_REF_JSON_DUMPS": "1",
+  "ATLAS_DFLASH_COPY_DRAFTS": "1",
+  "ATLAS_DFLASH_COPY_REPLY_MATCH": "16",
+  "ATLAS_DFLASH_COPY_MISS_MAX": "3"
+ },
+ "server_argv": [
+  "serve",
+  "${model}",
+  "--kernel-target=glm-5.3-flash",
+  "--world-size=2",
+  "--tp-size=2",
+  "--max-seq-len=1048576",
+  "--gpu-memory-utilization=0.88",
+  "--ep-size=2",
+  "--max-prefill-tokens=8192",
+  "--max-batch-size=4",
+  "--max-num-seqs=4",
+  "--oom-guard-mb=4096",
+  "--kv-cache-dtype=fp8_g128",
+  "--ssm-h-dtype=f32",
+  "--ssm-rollback-mode=records",
+  "--block-size=16",
+  "--ssm-cache-slots=16",
+  "--enable-prefix-caching",
+  "--ssm-checkpoint-interval=0",
+  "--gpu-ordinal=0",
+  "--swap-space-gb=0",
+  "--no-auto-swap",
+  "--no-tui",
+  "--lm-head-dtype=bf16",
+  "--request-timeout=0",
+  "--video-allow-ffmpeg",
+  "--vision-allow-remote-images",
+  "--tool-call-parser=poolside_v1",
+  "--video-max-frames=32",
+  "--dflash",
+  "--draft-model=${drafter}",
+  "--dflash-gamma=8",
+  "--dflash-window-size=2048",
+  "--mtp-gate=force"
+ ]
+}

install/profiles/4x512k.json

@@ -179,7 +179,10 @@
   "ATLAS_DFLASH_CTX_ASYNC_POS": "1",
   "ATLAS_GLM_STRICT_SPEC": "1",
   "ATLAS_DFLASH_CONF_WIDTH": "1",
-  "ATLAS_USE_HF_REF_JSON_DUMPS": "1"
+  "ATLAS_USE_HF_REF_JSON_DUMPS": "1",
+  "ATLAS_DFLASH_COPY_DRAFTS": "1",
+  "ATLAS_DFLASH_COPY_REPLY_MATCH": "16",
+  "ATLAS_DFLASH_COPY_MISS_MAX": "3"
  },
  "server_argv": [
   "serve",

install/profiles/8x128k.json

@@ -178,7 +178,10 @@
   "ATLAS_DFLASH_CTX_ASYNC_POS": "1",
   "ATLAS_GLM_STRICT_SPEC": "1",
   "ATLAS_DFLASH_CONF_WIDTH": "1",
-  "ATLAS_USE_HF_REF_JSON_DUMPS": "1"
+  "ATLAS_USE_HF_REF_JSON_DUMPS": "1",
+  "ATLAS_DFLASH_COPY_DRAFTS": "1",
+  "ATLAS_DFLASH_COPY_REPLY_MATCH": "16",
+  "ATLAS_DFLASH_COPY_MISS_MAX": "3"
  },
  "server_argv": [
   "serve",

install/serve.py

@@ -87,6 +87,25 @@ def disk_prefix_cache(gb):
             'ATLAS_SSM_TIER_SLOTS': '2'}
 
 
+def kv_shard(environ, disk_tier):
+    """Engine settings that split GLM's MLA latent KV between the two ranks,
+    when SPARKGLM_KV_SHARD=1 (default off): each rank stores half the blocks'
+    latents, so the shared pool holds about 1.8x the tokens for the same
+    memory (with the display carveout at 0.89: 120,186 blocks against 67,608;
+    single-request decode unchanged, four at once -3%). The disk prefix-cache
+    tier spills latents by physical block, which a shard cannot serve, so the
+    two exclude each other."""
+    switch = environ.get('SPARKGLM_KV_SHARD', '0')
+    if switch not in ('0', '1'):
+        raise ValueError('SPARKGLM_KV_SHARD must be 0 or 1')
+    if switch == '0':
+        return {}
+    if disk_tier:
+        raise ValueError('SPARKGLM_KV_SHARD=1 and the disk prefix cache (SPARKGLM_PREFIX_CACHE_GB) '
+                         'cannot be combined; switch one off')
+    return {'ATLAS_GLM_KV_SHARD': '1', 'ATLAS_GLM_KV_SHARD_COMPACT': '1'}
+
+
 def launch(environ, profile):
     rank = environ.get('NODE_RANK', '')
     if rank not in ('0', '1'):
@@ -125,7 +144,9 @@ def launch(environ, profile):
             raise ValueError('SPARKGLM_GPU_MEMORY_UTILIZATION must be 0.80 to 0.95')
         args = [f'--gpu-memory-utilization={util}' if a.startswith('--gpu-memory-utilization=') else a
                 for a in args]
-    env.update(disk_prefix_cache(environ.get('SPARKGLM_PREFIX_CACHE_GB', '')))
+    disk_tier = disk_prefix_cache(environ.get('SPARKGLM_PREFIX_CACHE_GB', ''))
+    env.update(disk_tier)
+    env.update(kv_shard(environ, disk_tier))
     name = environ.get('SERVED_MODEL_NAME', 'glm-5.3-flash-atlas')
     if not re.fullmatch(r'[A-Za-z0-9][A-Za-z0-9._/-]{0,127}', name):
         raise ValueError('Invalid SERVED_MODEL_NAME')