From d2fda4ef640b67bd4c4ca2ef80686fc0c42554d2 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=EA=B9=80=EB=B4=89=EB=AF=BC?=
 <102722487+bongmiin@users.noreply.github.com>
Date: Fri, 12 Jun 2026 16:39:01 +0900
Subject: [PATCH 1/2] Fix Fix Qwen3.5/3.6 load_weights stacked-mapping name
 mutation

The stacked_params_mapping loop in qwen3_5.py and qwen3_5_mtp.py mutates
name in place via name.replace(weight_name, param_name), then continues
the loop when the mapped name is not in params_dict. Because the mutated
name is not restored, a later mapping entry can re-apply .replace() to an
already-mapped name (e.g. up_proj is a substring of gate_up_proj), producing
invalid names such as gate_gate_up_proj and qkqkv_proj, so weights are not
loaded and FP8 post-processing fails.

Use a local name_mapped candidate and only assign it back to name after a
successful params_dict lookup, matching the safe pattern already used in the
expert branch.

Refs #474
---
 vllm/patches/vllm_for_multi_arc.patch | 14 ++++++++------
 1 file changed, 8 insertions(+), 6 deletions(-)

diff --git a/vllm/patches/vllm_for_multi_arc.patch b/vllm/patches/vllm_for_multi_arc.patch
index 9e305f1b..1f204fa5 100644
--- a/vllm/patches/vllm_for_multi_arc.patch
+++ b/vllm/patches/vllm_for_multi_arc.patch
@@ -11851,16 +11851,17 @@ index 000000000..b01b55f70
 +                if "mlp.experts" in name:
 +                    continue
 +
-+                name = name.replace(weight_name, param_name)
++                name_mapped = name.replace(weight_name, param_name
 +                # Skip loading extra bias for GPTQ models.
-+                if name.endswith(".bias") and name not in params_dict:
++                if name_mapped.endswith(".bias") and name_mapped not in params_dict:
 +                    continue
 +                # Skip layers on other devices.
 +                if is_pp_missing_parameter(name, self):
 +                    continue
 +                # name = apply_attn_prefix(name, params_dict)
-+                if name not in params_dict:
++                if name_mapped not in params_dict:
 +                    continue
++                name = name_mapped
 +                param = params_dict[name]
 +
 +                # Use weight_loader_v2 for tuple shard_id (GDN projections).
@@ -12604,15 +12605,16 @@ index 000000000..547c1885a
 +                if "mlp.experts" in name:
 +                    continue
 +
-+                name = name.replace(weight_name, param_name)
++                name_mapped = name.replace(weight_name, param_name)
 +                # Skip loading extra bias for GPTQ models.
-+                if name.endswith(".bias") and name not in params_dict:
++                if name_mapped.endswith(".bias") and name_mapped not in params_dict:
 +                    continue
 +                # Skip layers on other devices.
-+                if is_pp_missing_parameter(name, self):
++                if is_pp_missing_parameter(name_mapped, self):
 +                    continue
 +                if name not in params_dict:
 +                    continue
++                name = name_mapped
 +                param = params_dict[name]
 +                weight_loader = param.weight_loader
 +                weight_loader(param, loaded_weight, shard_id)

From cd4bd8e8eff056492bf3c84cd1689f7da72b73a8 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=EA=B9=80=EB=B4=89=EB=AF=BC?=
 <102722487+bongmiin@users.noreply.github.com>
Date: Fri, 12 Jun 2026 16:48:26 +0900
Subject: [PATCH 2/2] Fix indentation and correct name mapping logic

---
 vllm/patches/vllm_for_multi_arc.patch | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/vllm/patches/vllm_for_multi_arc.patch b/vllm/patches/vllm_for_multi_arc.patch
index 1f204fa5..89976a90 100644
--- a/vllm/patches/vllm_for_multi_arc.patch
+++ b/vllm/patches/vllm_for_multi_arc.patch
@@ -11851,7 +11851,7 @@ index 000000000..b01b55f70
 +                if "mlp.experts" in name:
 +                    continue
 +
-+                name_mapped = name.replace(weight_name, param_name
++                name_mapped = name.replace(weight_name, param_name)
 +                # Skip loading extra bias for GPTQ models.
 +                if name_mapped.endswith(".bias") and name_mapped not in params_dict:
 +                    continue