[yueming-yuan]

Stacked on #42820. The model- and format-specific engine fixes miles' p2p weight update needs; #42820 keeps the general engine support.

Derived state after writes that bypass a model's loader (moved from #42820, plus one gap the review found):

fix	symptom	cause
GemmaRMSNorm gemma_weight	Qwen3.5 serves the old norm after a p2p update; fused allreduce kernels read gemma_weight	the norm's weight loader derives gemma_weight = weight + 1, and p2p, remote-instance, sharded-state and dummy loads write the weight without it; post_load_weights now refreshes it
gemma_weight at every session end	a session whose direct or custom load wrote the norm weight kept the old gemma_weight	those loads bypass the norm's loader yet mark the session loaded, so end_weight_update skipped post_load_weights; it now refreshes gemma_weight either way
GLM-lite NextN post-load	the GLM-4.7-Flash draft skipped its NextN post-load after a p2p update	post_load_weights ran the target's path on the draft; the NextN draft now pins is_nextn=True

Quantized checkpoints. Four fixes found by running miles' p2p weight-update e2e with quantized rollout checkpoints on GB300 (fp8 block, mxfp8, nvfp4 on GLM-5.2 cut to 5 layers; fp8 on Qwen3.5-35B-A3B cut to 4 layers with its MTP draft). None is specific to p2p: each also hits sglang's own weight update or its startup. One commit each.

fix	symptom	cause
EAGLE draft KV budget for Qwen3.5	the draft's KV pool is allocated outside mem_fraction_static; OOM at startup on the 4-layer cut (127.6 GB draft pool, 82 GB left at 0.7)	_resolve_eagle_aux_hidden_state reads the draft depth from num_nextn_predict_layers; Qwen3.5 names it mtp_num_hidden_layers, so the target's per-token cost counted no draft layer
UE8M0 scales of non-128-multiple rows	RuntimeError: shape '[20, 128, 48]' is invalid for input of size 125952 in the weight checker on every GLM-5 fp8 update on Blackwell	_inverse_transform_scale_ue8m0_impl viewed the per-row packed scales as whole 128-row blocks; GLM-5's fused q_a + kv_a has 2624 rows
mxfp8 shards of the fused indexer	the fused bf16 wk_weights_proj off by up to 5e4 after loading an mxfp8 checkpoint with a quantized wk	_load_fused_indexer_wk dequantized with block_quant_dequant, multiplying by the uint8 UE8M0 exponents as integers
NVFP4 MoE input scales in the weight check	every NVFP4 MoE update fails the check on w13_input_scale / w2_input_scale only	they are calibrated activation scales no update carries (miles' quantizer emits weight, weight_scale, weight_scale_2); the checker reset them to random before the update. Now _skip_weight_check, as the KV cache scales

On the draft budget: with Qwen3.5-35B-A3B's 10 full-attention layers the unbudgeted draft is a tenth of the KV pool, which fits at 0.6–0.7 and so went unnoticed; the 4-layer cut makes it as large as the target's.

Tests

• unit/model_loader/test_default_model_loader.py: gemma_weight follows a norm weight written without its loader (CPU).
• unit/model_executor/model_runner_components/test_weight_updater_published_storage.py: a session whose load bypassed the norm's loader still refreshes gemma_weight (CPU); fails on its parent.
• unit/models/test_glm4_moe_lite_nextn_post_load.py: the GLM-lite NextN draft runs the NextN post-load (CPU).
• unit/model_executor/model_runner_components/test_spec_aux_hidden_state.py: an MTP depth named mtp_num_hidden_layers counts as the EAGLE draft's layers (CPU).
• unit/layers/quantization/test_fp8_utils.py: the UE8M0 round trip also covers GLM-5's (2624, 6144) (CUDA).
• unit/models/test_deepseek_fused_indexer_mxfp8.py: mxfp8 wk and weights_proj reach the fused param dequantized with their UE8M0 scales (CPU).
• unit/layers/quantization/test_modelopt_nvfp4_moe_weight_check.py: the checker skips the NVFP4 MoE input scales and still checks the weight scales (CPU).

On an H200 devbox (radixark/miles:dev), all changed test files of this branch and #42820 pass (35 passed, 1 skipped: an SM100-only case). On #42820's head each quantized case fails for its cause: None == 1, the [20, 128, 48] view error, the fused wk mismatch, w13_input_scale not skipped.

End to end on GB300 with the UE8M0 fix applied: GLM-5.2 5-layer fp8 compares every fp8 tensor within quantization ULP, fused_qkv_a_proj_with_mqa included. What remains there is on miles' side: its fp8 / mxfp8 converters quantize the interleaved indexer's wk, which sglang keeps in bf16 and the trainer sends in bf16.

---

CI States

Latest PR Test (Base): ❌ Run #38300076576
Latest PR Test (Extra): ❌ Run #38300075941
Latest PR Test (AMD ROCm 10): ❌ Run #38300076381