llama.cpp qwen4exp 'SPLIT_MODE_TENSOR not implemented': Update the Runtime
Fix the qwen4exp SPLIT_MODE_TENSOR not implemented error by updating stale llama.cpp builds, verifying the actual binary, and using layer split as a fallback.
Approximately 3 min read
If llama.cpp exits with:
LLAMA_SPLIT_MODE_TENSOR not implemented for architecture 'qwen4exp'
update the llama.cpp runtime that is actually running. This exact advice has changed: current llama.cpp no longer puts qwen4exp on the tensor-split architecture deny-list, and current upstream reports show Qwen3.8-Flash-Next running with --split-mode tensor.
If updating the real runtime is not immediately possible, switch to:
-sm layer
as the compatibility fallback. Do not spend time changing context, KV-cache precision, or -ts ratios to solve an explicit architecture capability error.
Why the old advice changed
Older qwen4exp builds could reject tensor split at model creation with the exact not implemented message. The current architecture gate in src/llama-arch.cpp returns false for a list of unsupported architectures; QWEN4EXP is not in that list.
The current multi-GPU documentation likewise says unsupported architectures fail with this message and recommends layer split, but qwen4exp is not listed among the unsupported families.
That means the same error today is strong evidence that the executable, container, GUI backend, Python binding, or packaged runtime is older than the source you are reading.
Verify the actual binary before rebuilding
Run:
command -v llama-server
readlink -f "$(command -v llama-server)"
llama-server --version
For a source build, also record:
git rev-parse HEAD
If a GUI or container launches llama.cpp, inspect the version inside that application or container. Updating a separate checkout does not replace a bundled backend.
-ts does not enable an unsupported split mode
These controls do different jobs:
-sm / --split-mode chooses layer, tensor, or another strategy
-ts / --tensor-split chooses device proportions
On an old build that rejects qwen4exp tensor mode, changing:
-ts 1,1
to another ratio cannot bypass the architecture gate. Update the runtime or use layer split first.
Current tensor split has its own requirements
Getting past the old architecture gate does not mean every tensor-split configuration is valid.
Current llama.cpp multi-GPU documentation requires Flash Attention for tensor mode and says quantized KV cache is not supported there. Check the exact options supported by your build and start from a minimal configuration.
For example, after updating:
llama-server \
-m /models/qwen3.8-flash-next.gguf \
-ngl all \
-sm tensor \
-ts 1,1 \
-fa on \
-ctk f16 \
-ctv f16
Treat the ratio as an example, not a universal optimum.
Do not confuse the old gate with SPLIT_AXIS_UNKNOWN
An older Qwen3.8-Flash-Next report, issue #27964, failed later with:
GGML_ASSERT(ret.axis != GGML_BACKEND_SPLIT_AXIS_UNKNOWN) failed
That is a different failure stage. The runtime accepted tensor mode, started constructing/distributing the graph, and then could not resolve a split state for an operation.
For that error, use the dedicated SPLIT_AXIS_UNKNOWN guide. Layer split remains a useful control, but do not describe a graph assertion as the same thing as an architecture gate.
How to verify the fix
After updating the runtime:
- confirm the version/commit from the executable that launches;
- start the same qwen4exp GGUF with tensor mode;
- verify the old
not implementedmessage is gone; - confirm context initialization completes;
- send a real prompt and check output;
- only then tune device ratios, context, cache types, and performance.
If the error changes to an allocator failure or SPLIT_AXIS_UNKNOWN, that is progress—the old capability gate is no longer the blocker, and the new error needs its own diagnosis.
Search-query answer
For the exact error:
LLAMA_SPLIT_MODE_TENSOR not implemented for architecture 'qwen4exp'
the current answer is:
update the actual llama.cpp runtime
-> retry tensor mode with current requirements
-> use layer split as a fallback if you cannot update
Do not rename the model, change quantization, or alter the article’s public URL to work around a stale runtime.