Hailo-10H M.2 (SKU 11): request LLM-friendly fitImage with shrunk integrated_nnc_dma_cma

Hardware: Seeed Hailo-10H M.2 (Board SKU-ID = 11), 2 GB H54G46CYRBX267 LPDDR4X

Firmware: HailoRT 5.3.0 / fitImage built 2026-03-23

Current behavior:

hailo-ollama LLM models all fail at create-LLM:

\[mmap_buffer.cpp:39\]  Failed to mmap buffer with errno:12

\[mmap_buffer.hpp:87\]  HAILO_OUT_OF_HOST_MEMORY(3)

hailortcli monitor: RAM 1194 MB total on device

Root cause (verified by DTB analysis):

SKU 11 kernel DTB inside fitImage reserves a 512 MB

`integrated_nnc_dma_cma,cma` no-map region at 0xb2000000.

Combined with phram (256 MB) + nnc_fw_cma (32 MB) + others,

~800 MB of the 2 GB physical LPDDR is permanently unavailable to

ARM Linux, leaving only 1194 MB. Every hailo-ollama HEF (1.3-2.3 GB)

exceeds this ceiling.

Verification (we tried):

Modified fitImage to shrink integrated_nnc_dma_cma to 128 MB,

patched outer FIT hash-1, deployed.

Driver loaded: customer_cert, scu_fw, u-boot-11.dtb.signed,

u-boot-spl, fitImage, image-fs all “programmed successfully”.

Then `vDMA transfer completed, triggering boot` followed by

`Timeout waiting for firmware file (-110)`.

Confirmed SPL/SCU rejects unsigned-or-resigned fitImage.

Rolled back to original fitImage; device fully restored.

Request:

Could Seeed (as Hailo customer key holder) provide an alternative

fitImage build for SKU 11 with `integrated_nnc_dma_cma,cma` size

reduced from 0x20000000 (512 MB) to e.g. 0x08000000 (128 MB)?

This is a single-line change in the SKU 11 kernel DTB inside the

FIT, signed with the customer key so the SoC accepts it.

We can also test intermediate sizes (192 MB, 256 MB) if 128 MB

is too tight for NNC LLM workload.

Hi @Rengmin_Wu,

The 2GB variant has limited capabilities to execute GenAI models - for those the 8GB variant is better suited.

You can still try to run the 1B parameters model Llama3.2-1B-Instruct: hailo_model_zoo_genai/docs/MODELS.rst at main · hailo-ai/hailo_model_zoo_genai · GitHub

May I ask what Hailo-10H M.2 you are using?

Thanks,