Inconsistent units ("M", " MiB", "M", " M")

[ericcurtin]

~$ docker model ls
MODEL NAME                                      PARAMETERS  QUANTIZATION    ARCHITECTURE  MODEL ID      CREATED       SIZE
hf.co/ggml-org/smolvlm-500m-instruct-gguf:q8_0  409M                        llama         982f124fa007  6 months ago  437M
hf.co/ggml-org/gemma-3-270m-gguf                268M                        gemma3        845da567e322  2 months ago  292M
smollm2                                         361.82 M    IQ2_XXS/Q4_K_M  llama         354bf30d0aa3  7 months ago  256.35 MiB
ericcurtin044/smollm2                           361.82 M    IQ2_XXS/Q4_K_M  llama         354bf30d0aa3  7 months ago  256.35 MiB
ignaciolopezluna020/embedding:dir-test          134.52 M    MOSTLY_Q2_K     llama         0a3c1f157564  10 days ago   82.41 MiB
embedding:dir-test                              134.52 M    MOSTLY_Q2_K     llama         4e6611fbb0e9  10 days ago   82.41 MiB
embedding                                       134.52 M    MOSTLY_Q2_K     llama         ae9c542496a3  10 days ago   82.41 MiB
ecurtin@LQ5W arm64 11:12:23 0
~$ docker model ls | awk '{print $6}'
MODEL
months
months
354bf30d0aa3
354bf30d0aa3
0a3c1f157564
4e6611fbb0e9
ae9c542496a3
~$ docker images
REPOSITORY                          TAG       IMAGE ID       CREATED       SIZE
alpine                              latest    4b7ce07002c6   3 weeks ago   13.3MB
rw4lll/openwebui-docker-extension   0.1.0     b7e59b9001cc   5 weeks ago   14.3MB

Makes it hard to use tools like awk, arrange things in columns, whitespace separated

[krissetto]

there should be a docker lib for these unit formatting scenarios: https://github.com/docker/go-units

[hwclass]

Hi! I'm new to this codebase but after @ericcurtin's encouragement, wanted to give this a shot. I've implemented a fix (which works on my machine :) ) for the inconsistent unit formatting issue. Would love to get feedback before opening a PR!

Approach

I've standardized unit formatting by removing spaces between numbers and units for both GGUF and Safetensors models.

For Safetensors (pkg/distribution/internal/safetensors/metadata.go):

• Removed spaces from formatParameters() unit arrays (" K" → "K", " M" → "M")
• Switched formatSize() to use units.BytesSize() instead of HumanSizeWithPrecision() + post-processing
  • BytesSize() outputs without spaces, uses binary units (MiB/GiB), and avoids scientific notation

For GGUF (pkg/distribution/internal/gguf/create.go):

• Added normalizeUnitString() using regex to remove spaces between numbers and units
• Applied to both Parameters and Size fields
• Handles all cases: "409M" (unchanged), "361.82 M" → "361.82M", "256.35 MiB" → "256.35MiB"

Tests: Updated expectations in both model_test.go files to match new format.

Why This Approach?

I used a hybrid approach because:

• Safetensors: We control formatting, so BytesSize() gives clean output directly
• GGUF: The parser already provides formatted strings, so regex normalization (pattern: ([0-9.]+)\s+([A-Za-z]+)) may be more efficient than re-parsing/reformatting (avoids type conversion errors, preserves original format)

After testing, BytesSize() is better than HumanSizeWithPrecision() because it has no spaces, no scientific notation, uses accurate binary units, and requires just one function call.

Results

Before:

PARAMETERS  SIZE
409M        437M
361.82 M    256.35 MiB
134.52 M    82.41 MiB

After:

PARAMETERS  SIZE
409M        437M
361.82M     256MiB
134.52M     82.41MiB

Now awk '{print $6}' works correctly for all models! 🎉

Files I've touched:

• pkg/distribution/internal/safetensors/metadata.go
• pkg/distribution/internal/gguf/create.go
• pkg/distribution/internal/safetensors/model_test.go
• pkg/distribution/internal/gguf/model_test.go

Since I'm still learning the codebase, I'd really appreciate any feedback on this approach. Does this align with the project's style? Any concerns or suggestions? If this looks good, I'll create a PR! Thanks!

[ericcurtin]

Sounds good @hwclass sometimes we lean on docker for influence:

$ docker images
REPOSITORY                          TAG       IMAGE ID       CREATED        SIZE
docker/model-runner                 latest    3840559d94a3   2 days ago     717MB
alpine                              latest    4b7ce07002c6   4 weeks ago    13.3MB
ubuntu                              latest    66460d557b25   5 weeks ago    139MB
ubuntu                              22.04     09506232a800   5 weeks ago    106MB
rw4lll/openwebui-docker-extension   0.1.0     b7e59b9001cc   6 weeks ago    14.3MB
testcontainers/ryuk                 0.13.0    31b31269d066   2 months ago   13.8MB
registry                            3         cd92709b4191   7 months ago   76.2MB

what would you think about MB no space like "docker images"?

Create the PR though for sure :)