Merge 0b829facda into 85fc35e8fa

This is using a different layers weight with .to(). Change it to use
the ops caster if the original layer is a comfy weight so that it picks
up dynamic_vram and async_offload functionality in full.

Co-authored-by: Rattus <rattus128@gmail.com>

comfy/text_encoders/ace15.py

@@ -3,6 +3,7 @@ import comfy.text_encoders.llama
 from comfy import sd1_clip
 import torch
 import math
+import comfy.utils
 
 
 def sample_manual_loop_no_classes(
@@ -42,6 +43,8 @@ def sample_manual_loop_no_classes(
     for x in range(model_config.num_hidden_layers):
         past_key_values.append((torch.empty([embeds.shape[0], model_config.num_key_value_heads, embeds.shape[1] + min_tokens, model_config.head_dim], device=device, dtype=execution_dtype), torch.empty([embeds.shape[0], model_config.num_key_value_heads, embeds.shape[1] + min_tokens, model_config.head_dim], device=device, dtype=execution_dtype), 0))
 
+    progress_bar = comfy.utils.ProgressBar(max_new_tokens)
+
     for step in range(max_new_tokens):
         outputs = model.transformer(None, attention_mask, embeds=embeds.to(execution_dtype), num_tokens=num_tokens, intermediate_output=None, dtype=execution_dtype, embeds_info=embeds_info, past_key_values=past_key_values)
         next_token_logits = model.transformer.logits(outputs[0])[:, -1]
@@ -54,8 +57,9 @@ def sample_manual_loop_no_classes(
         if eos_token_id is not None and eos_token_id < audio_start_id and min_tokens < step:
             eos_score = cfg_logits[:, eos_token_id].clone()
 
+        remove_logit_value = torch.finfo(cfg_logits.dtype).min
         # Only generate audio tokens
-        cfg_logits[:, :audio_start_id] = float('-inf')
+        cfg_logits[:, :audio_start_id] = remove_logit_value
 
         if eos_token_id is not None and eos_token_id < audio_start_id and min_tokens < step:
             cfg_logits[:, eos_token_id] = eos_score
@@ -63,7 +67,7 @@ def sample_manual_loop_no_classes(
         if top_k is not None and top_k > 0:
             top_k_vals, _ = torch.topk(cfg_logits, top_k)
             min_val = top_k_vals[..., -1, None]
-            cfg_logits[cfg_logits < min_val] = float('-inf')
+            cfg_logits[cfg_logits < min_val] = remove_logit_value
 
         if top_p is not None and top_p < 1.0:
             sorted_logits, sorted_indices = torch.sort(cfg_logits, descending=True)
@@ -72,7 +76,7 @@ def sample_manual_loop_no_classes(
             sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
             sorted_indices_to_remove[..., 0] = 0
             indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
-            cfg_logits[indices_to_remove] = float('-inf')
+            cfg_logits[indices_to_remove] = remove_logit_value
 
         if temperature > 0:
             cfg_logits = cfg_logits / temperature
@@ -90,6 +94,7 @@ def sample_manual_loop_no_classes(
         attention_mask = torch.cat([attention_mask, torch.ones((2, 1), device=device, dtype=attention_mask.dtype)], dim=1)
 
         output_audio_codes.append(token - audio_start_id)
+        progress_bar.update_absolute(step)
 
     return output_audio_codes
 

comfy/text_encoders/llama.py

@@ -6,6 +6,7 @@ import math
 
 from comfy.ldm.modules.attention import optimized_attention_for_device
 import comfy.model_management
+import comfy.ops
 import comfy.ldm.common_dit
 import comfy.clip_model
 
@@ -627,10 +628,10 @@ class Llama2_(nn.Module):
         mask = None
         if attention_mask is not None:
             mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, seq_len, attention_mask.shape[-1])
-            mask = mask.masked_fill(mask.to(torch.bool), float("-inf"))
+            mask = mask.masked_fill(mask.to(torch.bool), torch.finfo(x.dtype).min)
 
         if seq_len > 1:
-            causal_mask = torch.empty(past_len + seq_len, past_len + seq_len, dtype=x.dtype, device=x.device).fill_(float("-inf")).triu_(1)
+            causal_mask = torch.empty(past_len + seq_len, past_len + seq_len, dtype=x.dtype, device=x.device).fill_(torch.finfo(x.dtype).min).triu_(1)
             if mask is not None:
                 mask += causal_mask
             else:
@@ -794,7 +795,19 @@ class Qwen3_2B_ACE15_lm(BaseLlama, torch.nn.Module):
         self.dtype = dtype
 
     def logits(self, x):
-        return torch.nn.functional.linear(x[:, -1:], self.model.embed_tokens.weight.to(x), None)
+        input = x[:, -1:]
+        module = self.model.embed_tokens
+
+        offload_stream = None
+        if module.comfy_cast_weights:
+            weight, _, offload_stream = comfy.ops.cast_bias_weight(module, input, offloadable=True)
+        else:
+            weight = self.model.embed_tokens.weight.to(x)
+
+        x = torch.nn.functional.linear(input, weight, None)
+
+        comfy.ops.uncast_bias_weight(module, weight, None, offload_stream)
+        return x
 
 class Qwen3_4B(BaseLlama, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):

requirements.txt

@@ -1,4 +1,4 @@
-comfyui-frontend-package==1.37.11
+comfyui-frontend-package==1.38.13
 comfyui-workflow-templates==0.8.31
 comfyui-embedded-docs==0.4.0
 torch