Transformers to MLX Model Conversion

MLX-LM repo

In all circumstances refer to the following repo and branch unless overridden by the user.

GitHub repo: https://github.com/ml-explore/mlx-lm Branch: main

Workflow

Phase 0: Set up working environment

1. Verify uv is installed, use it to create a virtual environment. Always use uv pip install (not plain pip install) to install packages in the environment — this avoids resolver conflicts and is significantly faster.
2. git clone the mlx-repo from GitHub. ALWAYS USE the repo and branch specified in the MLX-LM repo section above.
3. Install mlx-lm in editable mode in the environment
4. Install transformers locally too, to ensure we have the latest version of the code: https://github.com/huggingface/transformers. You can clone with --depth 1 because it has many refs. Also install accelerate for device_map support.
5. Make sure huggingface_hub is installed too, to use the hf CLI for model download. It comes with the previous libraries as a dependency, so you may need to wait for it if you run stuff in parallel.
6. Use the venv binaries directly rather than activating: .venv/bin/python script.py, .venv/bin/hf download ..., .venv/bin/mlx_lm.generate .... This avoids source .venv/bin/activate (which executes a shell script that triggers permission prompts and doesn’t propagate to background commands). The CLI entry points and Python interpreter resolve their own environment from their installed location, so no activation is needed.
7. Discover and select target models, if the user did not provide a list of models to consider. Search the Hub for all models using the target architecture to convert, identify canonical variants across sizes, and confirm targets with the user. Use Model Discovery and Config Analysis to learn how.
8. Download a local copy of the model(s) from Hugging Face using the hf CLI. Save the models to local folders inside the project using --local-dir, this will allow you to read contents easily.

Phase 1: Analyze Source Models

1. Read config.json for model_type and architecture details
2. Verify native transformers support in .../transformers/models/<model_type>/, residing in the virtual environment
3. Read the transformers modeling_<model_type>.py implementation carefully
4. If the model is a vision-language model, convert the text-generation backbone, but inform the user in your report. Do not proceed if the model is not a language model at all (and explain why).
5. Analyze config variance across target models and cross-reference with conditional code paths in transformers. See Model Discovery and Config Analysis.
6. Pay special attention to unique or novel architecture variants the model may introduce. Highlight them in your report, and be vigilant for subtle bugs that may be difficult to detect in tests. Do some ad-hoc testing for them if necessary.

Some key config fields to check:

• rope_parameters, position_embedding_type - See section on RoPE later
• layer_types - For hybrid models (e.g., mamba + attention)
• Scalar multipliers: embedding_multiplier, attention_multiplier, residual_multiplier, logits_scaling
• torch_dtype or dtype or text_config.dtype — The model’s intended runtime precision. Some models (especially VL wrappers) nest this inside text_config rather than exposing it as top-level torch_dtype (or dtype). If the converter can’t find it, it skips the dtype casting step entirely, leaving stray float32 weights in the converted model. Verify which field the model uses and confirm mlx_lm.convert detects it. Use scripts/resolve_dtype.py to determine the expected dtype from the config (or the safetensors header as a fallback) and record the source — this becomes the basis for the dtype tests in Phase 4 and the manifest in Phase 8.

Phase 2: Find Reference MLX Models

Examples from mlx-lm/mlx_lm/models/:

• Standard attention: llama.py
• MLA (Multi-head Latent Attention): deepseek_v3.py — see MLA Reference
• MoE (Mixture of Experts): deepseek_v3.py, mixtral.py
• GatedDeltaNet (linear attention): qwen3_next.py, qwen3_5.py, olmo_hybrid.py
• Hybrid (Mamba + Attention): nemotron_h.py
• Sliding window: mistral.py

Start from the closest existing model rather than building from scratch — see Code Selection and Simplification.

Phase 3: Create MLX Model

Create mlx-lm/mlx_lm/models/<model_type>.py (self-contained single file):

@dataclass
class ModelArgs:
    # Config params with defaults
    position_embedding_type: str = "rope"  # Check for "nope"!

class Attention(nn.Module): ...
class MLP(nn.Module): ...
class DecoderLayer(nn.Module): ...
class LanguageModel(nn.Module): ...

class Model(nn.Module):
    def sanitize(self, weights): ...  # Convert weight names/shapes
    def shard(self, group): ...       # Distributed inference

    @property
    def cast_predicate(self):         # Keep specific weights in float32
        return lambda _, m, _: isinstance(m, MoEGate) and "bias" in m

sanitize() patterns — see Weight Sanitization for detailed guidance on how to make weights loading compatible with both transformers and MLX-converted weights. Here are some common examples:

• Stack experts: experts.{N}.gate_proj → switch_mlp.gate_proj shape (num_experts, ...)
• Remove: rotary_emb.inv_freq, extra layers (e.g., MTP heads)
• 3D MoE weights: May need reshape from (num_experts, hidden, intermediate)
• Shared/tied weights: Replicate from source blocks to all tied layers
• Different model sizes may use different weight formats — detect and branch
• MLA kv_b_proj splitting: dequantize → reshape by heads → split into embed_q/unembed_out — see MLA Reference

RoPE:

• rope_interleave=True → traditional=True in MLX
• position_embedding_type="nope" → Do NOT apply RoPE at all
• rope_parameters nested dict: some models store RoPE config as {"rope_theta": ..., "rope_type": ...} instead of top-level fields — extract in __post_init__. See MLA Reference.
• Null handling: rope_parameters can be null, absent, or a dict with null values (e.g., {"rope_theta": null}). Use .get() with defaults and test for None explicitly in __post_init__. Verify the different paths the transformers codebase uses to deal with all variants present in the config files.

Conversion and quantization: Use mlx_lm.convert to validate sanitize and produce quantized models — see Conversion and Quantization. Only do this after testing transformers weights, or if the model is too large to run unquantized.

Phase 4: Testing

Note: For very large models that can’t fit in fp16 / bf16, the standard comparison workflow isn’t feasible. Use the quantize-first strategy instead — see Large Models. For models that require multiple machines, see Distributed Inference.

Test 1: Generation comparison (do first - easier to spot garbage output)

Pre-validation: run mlx_lm.generate with a prompt and verify the output is coherent. This will use the default sampling method taken from the model’s generation_config.json.

Example:

mlx_lm.generate --model zai-org/GLM-4.7-Flash --prompt "Who are you?"

Use an appropriate prompt depending on the model capabilities (for example, is it a base model or does it use a chat template). If it’s a base model, use an inviting prefix sentence to be continued, not just a single word. If it’s a conversational model, ask an interesting question. If the model card mentions specific capabilities (for example, the model was trained for code generation), use a relevant topic.

If generated text is garbage, use the layer-by-layer tests to try to identify implementation mismatches.

For more controlled generation (greedy mode) and comparison with transformers, use the compare_generation.py script.

When using this script, make a judgement call about whether the generations are semantically similar. They will rarely match exactly, because of differences in the compute device, kernel implementations, and other reasons.

Do not declare success if generation is not reasonable - check for numerical and implementation differences and iterate until you get a good quality output, using the transformers one as the quality reference.

Test 2: Output dtype

Resolve the expected dtype with scripts/resolve_dtype.py — it tries config fields first, then falls back to the most common dtype in the safetensors header. Always record the source in the conversion report (e.g., config.dtype, safetensors header (BF16=292/293, F32=1/293)) so reviewers know the basis for the check.

python scripts/resolve_dtype.py /path/to/model
# dtype: bfloat16
# source: safetensors header (BF16=292/293, F32=1/293)

Then run a forward pass and verify the output dtype matches with scripts/check_dtype.py:

python scripts/check_dtype.py /path/to/model bfloat16
# {"output_dtype": "mlx.core.bfloat16", "expected_dtype": "bfloat16", "match": true}

Exits non-zero with a FAIL: message on stderr if the runtime dtype doesn’t match.

A single float32 weight (e.g., a norm weight promoted by v + 1.0 in sanitize) will silently propagate float32 through the entire model via dtype promotion, causing up to 5x slower inference on Apple Silicon.

The same script works after quantization — the runtime output dtype is unaffected by weight quantization (only the storage of matmul weights changes). Re-run it on the quantized model in Phase 5 to confirm.

Test 3: Numerical comparison

Use the compare_predictions script to compare logits from a single forward pass and extract some statistics. Use a prompt of at least 100 tokens (e.g., a paragraph of text) — short prompts (5-10 tokens) can hide positional encoding bugs because RoPE error grows with position. A wrong rope_theta may pass top-1 comparison at 5 tokens but diverge significantly at 100+.

Point out any major differences you observe. A max logits difference in ~0.5-1.0 (or larger, for models with many layers where errors accumulate) is usual between CPU transformers and Metal MLX. If you see suspicious values, run a layer-by-layer comparison even if the generation test seemed to pass.

If you run a top-5 or top-10 comparison, include the top-10 list in the GitHub PR you’ll open when you’re done (unless they are identical).

Test 4: Layer-by-layer (if predictions don’t match — and as a routine sanity check)

Use scripts/compare_layers.py. One forward pass per framework captures the output of every decoder layer, the post-final-norm state, and the logits, then compares them side-by-side with absolute and relative diffs.

python scripts/compare_layers.py /path/to/model "<prompt>"

Read the table from the bottom up:

• logits row: the most important. If logits agree, the implementation is end-to-end correct even if hidden states drift along the way.
• post-norm row: hidden state after the final RMSNorm. Should be small relative diff.
• layer N rows: per-layer outputs. Look for the first sharp jump in rel diff — that’s where divergence is introduced.

The script is architecture-agnostic: it wraps each entry in model.[language_model.]model.layers with a thin capture proxy (preserving the layer’s class so isinstance checks in hybrid models keep working). It does NOT compare the embedding output, since pre-layer transforms (multipliers, per-layer scaling) differ across architectures and the embedding row is rarely informative.

Healthy implementations show <1% relative diff at every position. A pattern where mid-layer divergence is relatively large but the logits row is small usually means the divergence cancels out at the head — annoying but not a correctness bug.

Include results in the report.

Test 5: Long sequence degradation

• Run a prompt that generates a long sequence, such as “Write an HTML and JavaScript page implementing space invaders”, or something appropriate depending on model capabilities. You can use mlx_lm.generate -m 16384 to allow up to 16K tokens to be generated.
• Observe if the output turns to garbage after a while. This usually signals mistakes in RoPE.

Phase 5: Quants

Use mlx_lm.convert to quantize the model. First, run scripts/resolve_dtype.py on the source checkpoint to confirm the storage dtype. If the model already deals with native quantized weights, skip this phase and inform the user. If it doesn’t, try a simple quantization to 4 bits and run generation tests to verify whether results are coherent.

After quantization, re-run scripts/check_dtype.py against the quantized model with the same expected dtype as the unquantized one. The runtime dtype shouldn’t change with quantization — only the storage of matmul weights does. A mismatch here usually means a stray float32 weight survived the quantize step.

Phase 6: MLX Tests

When appropriate, create tests for the model you just converted, following the structure and criteria used in the mlx-lm repo.

Phase 7: Report

Always include the following in your output report:

• The list of models and variants selected for conversion, and tested.
• Notable architecture decisions and novelties introduced.
• Non-trivial implementation details that are worth noting.
• The expected output dtype per variant and where it came from (e.g., config.dtype vs safetensors header). Use the results from scripts/resolve_dtype.py.
• At least a meaningful generation example of ~200 tokens or more.
• A summary of the numerical differences you found.
• Per-layer comparison results.
• Errors that you couldn’t solve and pointers to analyze them.
• If changes to mlx-lm were needed, explain in detail what they were, and provide links to the equivalent transformers implementation. This should be rare (bugs or missing features).
• Reflect upon learnings that we could consider to incorporate back into the Skill for future reference. Trivial discoveries that are easy for you to replicate in future sessions are not necessary; focus on initial error cases that required some iteration.

Please, output this report in your conversation, and also create a markdown file for reference.

Phase 8: GitHub PR

• Create a feature branch for your changes (e.g., add-<model_type>) — never commit directly to the main working branch. See GitHub PR Workflow for details.
• ALWAYS ask for confirmation before submitting a PR. The user may have questions or want to run additional tests, or include more variants.
• Make sure the gh cli is authenticated.
• Ensure you use the github repo specified at the beginning of this document (even if it’s not the canonical one).
• In the PR, include:
  • A suitable title. But please, prefix it with [transformers-to-mlx skill] for disclosure.
  • Also disclose that all tests and results were obtained with the skill.
  • Your summary report in the PR.
  • In addition to the summary report, include test commands and their outputs. For example, if you ran a long sequence generation with mlx_lm.generate, include the command and the full output. Also include the source code you used to compare numerical results, when appropriate. Same thing for the dtype tests and everything else.
• After the mlx-lm PR is created, generate a test manifest and open a PR to the test harness repo. See GitHub PR Workflow — Test Manifest for the format and procedure.

Critical Lessons

Garbage Output Despite Loading

Model loading and generating without crashes does NOT mean it works. Always verify:

1. Output is coherent text (not “lylyly Usesfinite”)
2. Top predictions match transformers
3. Logits statistics are in same ballpark

Position Embeddings

This is a usual source of problems.

Check position_embedding_type in config:

• "rope" - Apply rotary position embeddings
• "nope" - Do NOT apply any position embeddings (common in hybrid models)

RoPE config can vary across variants of the same architecture. For example, OLMo-Hybrid base/SFT use rope_theta=10000 while the DPO variant uses NoPE (no positional embeddings at all, via rope_parameters: {rope_theta: null}). Always diff RoPE-related config fields across ALL target checkpoints — base, SFT, DPO, and different sizes — before implementation.

# Conditional RoPE
use_rope = getattr(args, 'position_embedding_type', 'rope') != 'nope'
if use_rope:
    self.rope = initialize_rope(...)
else:
    self.rope = None

Scalar Multipliers

Some models (Granite, etc.) use scaling factors that MUST be applied exactly:

• embedding_multiplier - Scale embeddings before first layer
• attention_multiplier - Replace 1/sqrt(head_dim) in attention
• residual_multiplier - Scale residual connections
• logits_scaling - Scale final logits

Missing any of these causes completely wrong predictions.

Weight Key Names: Checkpoint vs Transformers Attributes

When building the MLX model, use the transformers model attribute names for your module definitions — not the raw checkpoint key names. These often differ: the checkpoint may use names like Wqkv or gate_up_proj while the transformers model class uses q_proj, k_proj, gate_proj, up_proj, etc.

The sanitize() method is where you remap checkpoint names to match the MLX model structure (which mirrors transformers attribute names). If the MLX model uses checkpoint names instead, weight loading will silently fail or require a convoluted sanitize.

To find the authoritative mapping, check conversion_mapping.py in the transformers model directory — see Weight Sanitization.

Numerical Precision

Some models define custom operations (e.g., gated RMSNorm, specialized activations) that are sensitive to precision. If the transformers code explicitly casts to float32, the MLX implementation must do the same:

class GatedRMSNorm(nn.Module):
    def __call__(self, x, residual):
        original_dtype = x.dtype
        x = x.astype(mx.float32)
        norm = mx.rsqrt(mx.mean(x * x, axis=-1, keepdims=True) + self.eps)
        x = (x * norm).astype(original_dtype) * self.weight
        return x

Missing a float32 cast may not cause obvious errors in short sequences but leads to degraded output quality (e.g., repetition) in longer generations. When reading the transformers source, watch for .float(), .to(torch.float32), or similar casts — they are there for a reason.

However, do check for MLX ops that may be internally upcasting in their implementation.

Dtype Contamination After Conversion

A single float32 parameter in a non-quantized layer (norms, biases, gating weights) can silently poison the entire forward pass through MLX’s dtype promotion rules. This causes no correctness errors — the model generates coherent text — but inference speed degrades dramatically (up to 5x slower) because all quantized matmuls run in float32 mode instead of bfloat16.

Common causes:

• Sanitize operations like v + 1.0 promote bfloat16 norm weights to float32 (Python’s 1.0 is float64/float32)
• Parameters initialized with mx.ones() or mx.zeros() default to float32 and persist if the checkpoint doesn’t include them

The contamination path: float32 norm weight → rms_norm outputs float32 → quantized matmul receives float32 input → outputs float32 → residual addition promotes hidden state to float32 → every subsequent layer inherits float32.

Always verify output dtype after conversion and after quantization (see Phase 5).

Hybrid and SSM Architectures

For models mixing layer types (e.g., Mamba + Attention) — see Hybrid and SSM Models for detailed guidance:

1. Check layer_types config for per-layer architecture
2. Each layer may have MULTIPLE blocks (not just one):
   • Block 1: Mamba OR Attention
   • Block 2: MoE + shared_mlp (for ALL layers)
3. Both Mamba and Attention layers often share the same MoE/MLP block
4. make_cache() must return the right cache type per layer (e.g., CacheList(KVCache(), ArraysCache()) for hybrid layers)
5. CUDA-only dependencies (mamba_ssm, causal_conv1d) won’t install on macOS — transformers uses fallback implementations

3D MoE Weights

Expert weights may be stored as 3D tensors:

• Shape: (num_experts, hidden_size, intermediate_size)
• Need proper reshape in sanitize()
• SwitchGLU expects specific shapes

Debugging Checklist

When predictions don’t match:

1. Embeddings - Should match exactly (mean, std)
2. After embedding scaling - Check multiplier applied
3. Layer-by-layer - Find first divergence point
4. Check config fields - position_embedding_type, multipliers
5. Layer structure - All blocks applied? (MoE often missed)
6. Attention scaling - Verify the scale factor matches transformers exactly (some models use non-standard formulas like (head_dim // 2) ** -0.5)
7. Precision casts - Check for .float() / .to(torch.float32) in transformers code, especially in norms and activations (see Numerical Precision above)
8. Repetitive output - If the model generates repetitive or looping text, this may indicate a shape or transpose error in the attention or linear attention layers (e.g., head dimensions swapped, wrong reshape order), not a sampling issue

Large Models

Models that consume a significant fraction of available RAM require special handling:

• Set sudo sysctl iogpu.wired_limit_mb=<value> to increase the wired memory limit
• Set MLX_METAL_FAST_SYNCH=1 for faster GPU synchronization
• mlx_lm.generate may OOM due to mx.async_eval double-buffering — use a manual generation loop
• Very large MoE models suffer memory system thrashing even with selective gather_qmm
• For models too large for one machine, use tensor parallelism via mlx.launch

See Large Models and Distributed Inference for detailed guidance.

Changes to Shared mlx-lm Code

Some conversions require modifications to shared mlx-lm infrastructure (ssm.py, Metal kernels, cache.py), not just a new model file. These changes affect all models using the same code. See Common Infrastructure Changes for guidance on localizing fixes and avoiding regressions.

General Guidance and Common Pitfalls

• Minimize code comments. Do not add comments unless the code would be legitimately confusing without one (which should be very rare). Comments must be reviewed just like code, so every unnecessary comment increases review burden. The PR description and tests provide all the context a reviewer needs.
• Greedy MLX decoding: sampler=lambda x: mx.argmax(x, axis=-1), not temperature=0
• Don’t override eos_token_id - use generation_config
• Use separate tokenizers for transformers/MLX tests
• apply_chat_template: use tokenize=True, return_tensors="pt" directly
• MPS limitations: use CPU for comparison tests
• Hybrid models: RoPE is often disabled (position_embedding_type: "nope")
• Every layer may need MoE block, not just attention layers

Output

Upon success, please save the mlx implementation in the appropriate place inside mlx-lm/mlx_lm/models, and generate a summary report with the results from the tests.

On failure (only declare when you run out of ideas after multiple iterations), please report your analysis as well as the results from tests you run.

List of scripts

Bundled scripts in scripts/ (relative to this skill):

Recommended order for the comparison flow:

1. compare_generation.py — quick sanity check (is output coherent?)
2. compare_predictions.py — numerical analysis (do predictions match?)
3. compare_layers.py — find divergence (where does it break?)

Usage:

# Resolve and verify dtype
python scripts/resolve_dtype.py /path/to/model
python scripts/check_dtype.py /path/to/model bfloat16

# Full generation comparison
python scripts/compare_generation.py /path/to/model --message "What is 2+2?"

# Numerical comparison with tolerance analysis
python scripts/compare_predictions.py /path/to/model "The quick brown fox..."

# Per-layer + logits comparison
python scripts/compare_layers.py /path/to/model "The quick brown fox..."

References