From 8f9de8ac1e7fede6c2a07b151ba5bf28e59d7943 Mon Sep 17 00:00:00 2001
From: Dmitrii <dmitrii.timoshenko16@gmail.com>
Date: Sat, 19 Sep 2026 19:31:59 +0200
Subject: [PATCH] fix(qwen35): write the bf16 prefill copies of
 ssm_alpha/beta_proj

The dense Qwen3.5 converter stopped emitting
`model.layers.{bid}.linear_attn.ssm_{alpha,beta}_proj.bf16.weight` in
d1d5232d: the three lines that store the dequantized copy were commented
out (with a leftover `sdfsdf` marker), while the comment added above in
the same commit still states that the bf16 copy "is derived from the
quantized tensor below".

Both `configs/qwen3.5_*.json` map those names as
`linear_{alpha,beta}_proj_prefill`, and the official FastFlowLM weights
contain them, so a model converted from a stock GGUF is missing 48
tensors and the runtime aborts with
`Weight not found: model.layers.0.linear_attn.ssm_alpha_proj.bf16.weight`.

Restore the block from `q4nx/models/qwen35_moe.py`, which the same commit
added with this code intact, including the `len(unpacked) == 3` guard that
keeps a non-quantized (BF16) source from raising
`ValueError: not enough values to unpack`.

Verified on Qwen3.5-4B: the converted file now carries all 474 tensors,
and the 48 restored bf16 copies are byte-identical to the official
Qwen3.5-4B-NPU2 weights.
---
 q4nx/models/qwen35.py | 49 +++++++++++++++++++++++++------------------
 1 file changed, 29 insertions(+), 20 deletions(-)

diff --git a/q4nx/models/qwen35.py b/q4nx/models/qwen35.py
index 30e92b4..57ca855 100644
--- a/q4nx/models/qwen35.py
+++ b/q4nx/models/qwen35.py
@@ -119,28 +119,37 @@ def convert(self, q4nx_path: str, weights_type: str = 'language'):
 
                     if ("ssm_alpha_proj" in self.forward_name_map[gguf_tensor.name] or "ssm_beta_proj" in self.forward_name_map[gguf_tensor.name]) \
                             and "bf16" not in  self.forward_name_map[gguf_tensor.name]: # for ssm_alpha_proj, the order is special
-               
-                        d, m, qw = unpacked
-                        # save bf16 copy
-                        w = gguf_tensor.dequantize()
-                        if reorder_linear_required:
-                            w = rearrange(w, '(q g) c -> (g q) c', q = 2).contiguous()
-
-                        # new_name = self.forward_name_map[gguf_tensor.name]
-                        # new_name = new_name.replace("alpha_proj", "alpha_proj.bf16").replace("beta_proj", "beta_proj.bf16")
-                        # self.q4nx_tensors[new_name] = w  sdfsdf
-                        if reorder_linear_required:
-                            print(f"[INFO] Reorder for {self.forward_name_map[gguf_tensor.name]}")
-                            d = rearrange(d, '(q g) c -> (g q) c', q = 2).contiguous()
-                            m = rearrange(m, '(q g) c -> (g q) c', q = 2).contiguous()
-                            qw = rearrange(qw, '(q g) c -> (g q) c', q = 2).contiguous()
 
-                        if (d.shape[0] < 32): # duplicate on the first dimension for better performance when the state size is small
-                            d = repeat(d, 'd c -> (r d) c', r = 2).contiguous()
-                            m = repeat(m, 'd c -> (r d) c', r = 2).contiguous()
-                            qw = repeat(qw, 'd c -> (r d) c', r = 2).contiguous()
+                        if len(unpacked) == 3:
+                            d, m, qw = unpacked
+                            # save bf16 copy
+                            w = gguf_tensor.dequantize()
+                            if reorder_linear_required:
+                                w = rearrange(w, '(q g) c -> (g q) c', q = 2).contiguous()
+
+                            new_name = self.forward_name_map[gguf_tensor.name]
+                            new_name = new_name.replace("alpha_proj", "alpha_proj.bf16").replace("beta_proj", "beta_proj.bf16")
+                            self.q4nx_tensors[new_name] = w
+                            if reorder_linear_required:
+                                print(f"[INFO] Reorder for {self.forward_name_map[gguf_tensor.name]}")
+                                d = rearrange(d, '(q g) c -> (g q) c', q = 2).contiguous()
+                                m = rearrange(m, '(q g) c -> (g q) c', q = 2).contiguous()
+                                qw = rearrange(qw, '(q g) c -> (g q) c', q = 2).contiguous()
+
+                            if (d.shape[0] < 32): # duplicate on the first dimension for better performance when the state size is small
+                                d = repeat(d, 'd c -> (r d) c', r = 2).contiguous()
+                                m = repeat(m, 'd c -> (r d) c', r = 2).contiguous()
+                                qw = repeat(qw, 'd c -> (r d) c', r = 2).contiguous()
 
-                        unpacked = (d, m, qw)
+                            unpacked = (d, m, qw)
+                        else:
+                            # save in bf16 into the file
+                            w = unpacked[0].contiguous().to(torch.bfloat16)
+                            if reorder_linear_required:
+                                w = rearrange(w, '(q g) c -> (g q) c', q = 2).contiguous()
+                            # save matrix in colum major order instead
+                            self.q4nx_tensors[self.forward_name_map[gguf_tensor.name]] = w.clone().transpose(0,1).contiguous()
+                            continue
 
 
                     if "ssm_conv1d" in self.forward_name_map[gguf_tensor.name]: