[NPU][Bug] check_mtp_only_grad fails on GLM-4.7-Flash

[Fulin-Gao]

check_mtp_only_grad fails on GLM-4.7-Flash

Symptom

tests/test_glm4.7_30B_A3B_npu.py (CI build) fails at Step 0 on the isolation assertion
assert len(non_mtp_nonzero_grads) == 0 in vime-ascend/vime/backends/megatron_utils/ci_utils.py:

AssertionError: Expected all non-MTP parameters to have zero gradients, but found 1302 with non-zero gradients.

Step 0 gradient stats:

[CI MTP Grad Check] Step 0: MTP params with non-zero grad: 32,
non-MTP params with non-zero grad: 1302
[CI MTP Grad Check] Non-MTP param with non-zero grad:
  module.module.embedding.word_embeddings.weight,        max_grad=1.02e-03   ← largest
  module.module.decoder.layers.0.input_layernorm.weight, max_grad=1.00e-04
  module.module.decoder.layers.0.self_attention.linear_proj.weight,                   max_grad=9.5e-06
  module.module.decoder.layers.0.self_attention.linear_q_down_proj.weight,             max_grad=9.6e-06
  module.module.decoder.layers.0.self_attention.linear_q_up_proj.layer_norm_weight,   max_grad=1.87e-05

Root Cause

check_mtp_only_grad runs unconditionally every step, but its assertion len(non_mtp_nonzero_grads) == 0 only holds under the docstring's "all outputs truncated" precondition. When rewards differ across samples in a GRPO group, advantages are non-zero, the policy loss is non-zero, and gradients flow to non-MTP params (embedding, decoder layers, ...), so the assertion fails.

Reproduction

Test entry: tests/test_glm4.7_30B_A3B_npu.py (16×NPU, ray job submit).
Key args:

--ci-test 
--enable-mtp-training 
--rollout-max-response-len 2048

Impact / Suggestions

• Affected
  The assertion assert len(non_mtp_nonzero_grads) == 0 in ci_utils.py::check_mtp_only_grad produces a false positive whenever rewards differ within a GRPO group.
• Candidate directions
  Reduce the max output length (e.g. lower --rollout-max-response-len) so all rollout responses truncate, satisfying the assertion's all-truncated precondition.