fix chatglm3 npu output (#11590)

python/llm/src/ipex_llm/transformers/npu_models/chatglm.py

@@ -64,7 +64,16 @@ def chatglm2_model_forward(
         rotary_pos_emb = rotary_pos_emb[position_ids]
     else:
         rotary_pos_emb = rotary_pos_emb[None, :seq_length]
-    rotary_pos_emb = rotary_pos_emb.transpose(0, 1).contiguous()
+    # ipex-llm change start: change rope cache shape
+    # rotary_pos_emb: [bsz, seq_len, rot_dim//2, 2]
+    cos, sin = rotary_pos_emb.permute(3, 0, 1, 2).chunk(2, dim=0)
+    cos = cos.squeeze(0).unsqueeze(1)
+    sin = sin.squeeze(0).unsqueeze(1)
+    cos = cos.repeat_interleave(2, dim=-1)
+    sin = sin.repeat_interleave(2, dim=-1)
+    # cos, sin: [bsz, 1, seq_len, rot_dim]
+    rotary_pos_emb = (cos, sin)
+    # ipex-llm change end
 
     # ipex-llm changes begin:
     # generate `causal_mask` and replace `full_attention_mask` with it
@@ -76,14 +85,6 @@ def chatglm2_model_forward(
                                   dtype=inputs_embeds.dtype, device=inputs_embeds.device)
         mask_value = torch.finfo(inputs_embeds.dtype).min
         causal_mask.masked_fill_(full_attention_mask, mask_value)
-    elif self.training or (inputs_embeds.device.type != "xpu" and past_key_values is None):
-        full_attention_mask = self.get_masks(input_ids,
-                                             past_key_values,
-                                             padding_mask=attention_mask)
-        causal_mask = torch.zeros([batch_size, 1, seq_length, full_attention_mask.size(-1)],
-                                  dtype=inputs_embeds.dtype, device=inputs_embeds.device)
-        mask_value = torch.finfo(inputs_embeds.dtype).min
-        causal_mask.masked_fill_(full_attention_mask, mask_value)
     else:
         causal_mask = None
 
@@ -174,24 +175,20 @@ def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
 
 
 @torch.jit.script
-def apply_rotary_pos_emb(x: torch.Tensor, rope_cache: torch.Tensor) -> torch.Tensor:
-    # x: [sq, b, np, hn]
-    sq, b, np, hn = x.size(0), x.size(1), x.size(2), x.size(3)
-    rot_dim = rope_cache.shape[-2] * 2
+def rotate_every_two(x: torch.Tensor):
+    x1 = x[:, :, :, ::2]
+    x2 = x[:, :, :, 1::2]
+    x = torch.stack((-x2, x1), dim=-1)
+    return x.flatten(-2)
+
+
+def apply_rotary_pos_emb(x: torch.Tensor, rope_cache: Tuple[torch.Tensor]) -> torch.Tensor:
+    # x: [bsz, n_head, seq_len, head_dim]
+    cos, sin = rope_cache
+    rot_dim = cos.size(-1)
     x, x_pass = x[..., :rot_dim], x[..., rot_dim:]
-    # truncate to support variable sizes
-    rope_cache = rope_cache[:sq]
-    xshaped = x.reshape(sq, -1, np, rot_dim // 2, 2)
-    rope_cache = rope_cache.view(sq, -1, 1, xshaped.size(3), 2)
-    x_out2 = torch.stack(
-        [
-            xshaped[..., 0] * rope_cache[..., 0] - xshaped[..., 1] * rope_cache[..., 1],
-            xshaped[..., 1] * rope_cache[..., 0] + xshaped[..., 0] * rope_cache[..., 1],
-        ],
-        -1,
-    )
-    x_out2 = x_out2.flatten(3)
-    return torch.cat((x_out2, x_pass), dim=-1)
+    x_out = x * cos + rotate_every_two(x) * sin
+    return torch.cat([x_out, x_pass], dim=-1)
 
 
 def chatglm2_attention_forward(
@@ -246,7 +243,7 @@ def chatglm2_attention_forward(
             key_states,
             value_states,
             attn_mask=attention_mask,
-            is_causal=q_len > 1 and bsz == 1,
+            is_causal=attention_mask is None and q_len > 1 and bsz == 1,
         )
         attn_weights = None
     else: