Feature or Model Request

Problem

The current MoE sparse_matmul implementation (layers/moe.py) and the underlying megablox/ops.py conflate two orthogonal concerns:

1. Quantization orchestration — obtaining QWIX rules, quantizing lhs/rhs tensors
2. Kernel backend dispatch — choosing between megablox Pallas, tokamax Pallas, or jax.lax.ragged_dot

This leads to confusing control flow where mblx.gmm() acts as both a "quantization-aware GMM wrapper" and a "backend router" via the
use_tokamax_backend flag. The megablox module shouldn't need to know about tokamax at all.

Current call paths in sparse_matmul (simplified)

if use_tokamax_gmm:
    if quantization:
        mblx.gmm(..., use_tokamax_backend=True)   # megablox wraps tokamax?
    else:
        tokamax.ragged_dot(...)                     # direct tokamax call
elif megablox:
    mblx.gmm(..., use_tokamax_backend=False)        # megablox native
else:
    jax.lax.ragged_dot(...)                         # JAX fallback

Issues:

• mblx.gmm imports and calls tokamax_backend internally — megablox shouldn't depend on tokamax
• Quantization logic is buried inside ops.py fwd/bwd functions (qpl.get_current_rule, qpl.quantize) rather than being a composable layer
• 4 code paths in sparse_matmul with subtle differences in tiling, quantization, and dtype handling
• Adding a new backend (e.g., Mosaic, Triton) requires modifying megablox internals

Proposed Direction

Separate the three concerns into composable layers:

1. Backend interface — a common GmmBackend protocol with gmm() / tgmm() methods, implemented by megablox, tokamax, and ragged_dot backends
   independently
2. Quantization wrapper — a standalone layer that handles qpl.get_current_rule / qpl.quantize and delegates the actual matmul to any backend
3. sparse_matmul — selects backend based on config, wraps it with quantization if needed, calls it

Desired flow (conceptual)

backend = get_backend(config) # tokamax / megablox / ragged_dot
if quantization:
backend = QuantizedGmm(backend) # wraps any backend with FP8 quantization
output = backend.gmm(inputs, kernel, group_sizes, tiling)

This would:

• Remove tokamax imports from megablox
• Make quantization composable with any backend
• Simplify sparse_matmul from 4 branches to ~2 lines
• Make adding new backends trivial

Additional Context

No response