From 70c91f94a7745a19d210a34d361daa3cd060f88e Mon Sep 17 00:00:00 2001
From: ent <bleysg@gmail.com>
Date: Wed, 24 Jun 2026 13:10:36 +1000
Subject: [PATCH] docs: GB10 memory is 128 GB (119 GiB), not 128 GiB

---
 README.md        | 8 ++++----
 docs/FINDINGS.md | 2 +-
 install.sh       | 4 ++--
 3 files changed, 7 insertions(+), 7 deletions(-)

diff --git a/README.md b/README.md
index 7846449..c358d18 100644
--- a/README.md
+++ b/README.md
@@ -2,7 +2,7 @@
 
 [`Qwen3.5-122B-A10B`](https://huggingface.co/Intel/Qwen3.5-122B-A10B-int4-AutoRound)
 (hybrid GDN + mamba + 128-expert MoE, ~10B active) running on a single
-**NVIDIA DGX Spark** (GB10 / SM121, 128 GiB unified) under **vLLM**, with
+**NVIDIA DGX Spark** (GB10 / SM121, 128 GB / 119 GiB unified) under **vLLM**, with
 **[DFlash](https://modal.com/blog/spec-is-all-u-need) block-diffusion speculative
 decode** and an optional **dense-bandwidth patch stack** — measured end-to-end,
 with a per-token bandwidth model that explains every number.
@@ -25,7 +25,7 @@ so it's a lever for *base / low-acceptance* serving, not for the agent path.
 - **Engine:** [`vLLM`](https://github.com/vllm-project/vllm) 0.23, sm121 build with the DFlash PRs, via the prebuilt image `ghcr.io/aeon-7/aeon-vllm-ultimate:2026-06-18-v0.23.0-dflashfix`. No host build — the four runtime patches in [`runtime/`](runtime/) are applied at serve time.
 - **Target:** [`Intel/Qwen3.5-122B-A10B-int4-AutoRound`](https://huggingface.co/Intel/Qwen3.5-122B-A10B-int4-AutoRound) — INT4 (AutoRound/GPTQ) routed experts + attention, BF16 shared experts/embeddings/head, ~62 GiB. (Safetensors, *not* GGUF — vLLM serves HF checkpoints directly.)
 - **Drafter:** [`z-lab/Qwen3.5-122B-A10B-DFlash`](https://huggingface.co/z-lab/Qwen3.5-122B-A10B-DFlash) — 0.8B / 6-layer non-causal block-diffusion drafter (block 16), shares the target's `embed_tokens` + `lm_head`, ~1.6 GiB.
-- **Hardware:** NVIDIA DGX Spark, GB10, SM121, 128 GiB LPDDR5X unified, ~273 GB/s.
+- **Hardware:** NVIDIA DGX Spark, GB10, SM121, 128 GB LPDDR5X unified (~119 GiB usable), ~273 GB/s.
 
 ## Quick start
 
@@ -58,12 +58,12 @@ Preview without running: `... | bash -s -- --help`.
 
 | | |
 |---|---|
-| Validated on | NVIDIA DGX Spark (GB10, SM121, 128 GiB unified) |
+| Validated on | NVIDIA DGX Spark (GB10, SM121, 128 GB / 119 GiB unified) |
 | Likely to work | other Blackwell with `--force` (untested) |
 | Runtime | Docker + NVIDIA container runtime (`docker run --gpus all`) |
 | Disk | ≥ 75 GiB free (image + weights); ≥ 150 GiB if `--build-hybrid` |
 | OS | aarch64 Linux (Grace) |
-| Memory | 128 GiB unified is enough for the model + DFlash drafter + KV @ 16k |
+| Memory | 128 GB / 119 GiB unified is enough for the model + DFlash drafter + KV @ 16k |
 
 GB10 is detected via `nvidia-smi --query-gpu=compute_cap` returning `12.1`;
 anything else needs `--force`.
diff --git a/docs/FINDINGS.md b/docs/FINDINGS.md
index 856fed3..c552346 100644
--- a/docs/FINDINGS.md
+++ b/docs/FINDINGS.md
@@ -1,7 +1,7 @@
 # FINDINGS — DFlash + dense levers for Qwen3.5-122B-A10B on DGX Spark
 
 Single-stream (c=1) decode of `Qwen3.5-122B-A10B` (hybrid GDN + mamba + 128-expert
-MoE, ~10B active) on GB10 / SM121, 128 GiB unified, ~273 GB/s. The agent this
+MoE, ~10B active) on GB10 / SM121, 128 GB / 119 GiB unified, ~273 GB/s. The agent this
 backs (Hermes) is ~73 % tool-calls. All numbers temperature 0.
 
 > **Credit.** This builds on [albond's recipe](https://github.com/albond/DGX_Spark_Qwen3.5-122B-A10B-AR-INT4)
diff --git a/install.sh b/install.sh
index 1b72421..1b6a87f 100755
--- a/install.sh
+++ b/install.sh
@@ -1,6 +1,6 @@
 #!/usr/bin/env bash
 # install.sh — Qwen3.5-122B-A10B + DFlash speculative decode on NVIDIA DGX Spark
-#              (GB10 / SM121, 128 GiB unified), via vLLM in Docker.
+#              (GB10 / SM121, 128 GB / 119 GiB unified), via vLLM in Docker.
 #
 #   curl -sSL https://raw.githubusercontent.com/Entrpi/qwen3.5-122B-A10B-on-spark/main/install.sh | bash
 #   curl -sSL https://raw.githubusercontent.com/Entrpi/qwen3.5-122B-A10B-on-spark/main/install.sh | bash -s -- --help
@@ -92,7 +92,7 @@ Flags:
   --nspec N               num_speculative_tokens (default 12 dflash/dense, 2 mtp, 0 base).
   --port N                Server port (default: $PORT).
   --ctx N                 max-model-len (default: $CTX).
-  --gpu-mem F             gpu-memory-utilization, keep <=0.84 on 128GiB (default: $GPU_MEM).
+  --gpu-mem F             gpu-memory-utilization, keep <=0.84 on 128GB (default: $GPU_MEM).
   --force                 Skip the GB10/SM121 host check.
   --no-smoke              Start the server but skip the Paris smoke test.