New PR for GPTOSS decode-only model (#603)

Signed-off-by: vbaddi <quic_vbaddi@quicinc.com>
Signed-off-by: Onkar Chougule <ochougul@qti.qualcomm.com>
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Signed-off-by: Mamta Singh <168400541+quic-mamta@users.noreply.github.com>
Co-authored-by: Vinayak Baddi <quic_vbaddi@quicinc.com>
Co-authored-by: Vinayak Baddi <vbaddi@qti.qualcomm.com>
Co-authored-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Co-authored-by: Mamta Singh <168400541+quic-mamta@users.noreply.github.com>
Signed-off-by: Vahid Janfaza <vjanfaza@qti.qualcomm.com>

Author: 5 people

QEfficient/base/pytorch_transforms.py

@@ -120,61 +120,109 @@ def apply(cls, model: nn.Module) -> Tuple[nn.Module, bool]:
 
 class SplitGateUpWeightsTransform(PytorchTransform):
     """
-    split fused Gate+Up weights and copy into the model
+    Split fused Gate+Up weights and copy into the model.
+    Handles both standard MoE models and GptOss models.
 
     For every transformer layer inside `model`:
-      • expects   <PREFIX>.experts.gate_up_proj   in the *source* `sd`
-      • copies halves into
-            <PREFIX>.experts.gate_proj     <-- Gate   [E,H,I]
-            <PREFIX>.experts.up_proj       <-- Up     [E,H,I]
+    • expects <PREFIX>.experts.gate_up_proj in the *source* `sd`
+    • copies halves into
+      <PREFIX>.experts.gate_proj     <-- Gate   [E,H,I]
+      <PREFIX>.experts.up_proj       <-- Up     [E,H,I]
+
+    Handles both interleaved weights (GptOss) and concatenated weights (standard MoE).
+    Also handles bias terms when present.
     """
 
     @classmethod
     def apply(cls, model: nn.Module) -> Tuple[nn.Module, bool]:
         transformed = False
         model_class = model.__class__.__name__ if hasattr(model, "model") else model.__class__.__name__
-
         if model_class not in VLM_SPLIT_GATE_UP_WEIGHTS:
             return model, transformed
 
         model_tmp = model.language_model if hasattr(model, "language_model") else model
-
         num_layers = len(model_tmp.model.layers)
         delete_fused_key = True
         sd = model_tmp.state_dict()
+
         for layer_idx in range(num_layers):
+            # Determine if this is a GptOss model or standard MoE model
+            is_gpt_oss = hasattr(model_tmp.model.layers[layer_idx], "mlp")
+
             # ---- build the textual prefix once per layer ----------
-            prefix = f"model.layers.{layer_idx}.feed_forward.experts."
+            if is_gpt_oss:
+                prefix = f"model.layers.{layer_idx}.mlp.experts."
+                experts = model_tmp.model.layers[layer_idx].mlp.experts
+            else:
+                prefix = f"model.layers.{layer_idx}.feed_forward.experts."
+                experts = model_tmp.model.layers[layer_idx].feed_forward.experts
 
             fused_key = prefix + "gate_up_proj"
             gate_key = prefix + "gate_proj"
             up_key = prefix + "up_proj"
 
-            # ---- split  [E,H,2I] → two  [E,H,I]  tensors ----------------------
-            fused = sd[fused_key]  # [E, H, 2I]  (no .weight here)
+            # Check if we have bias terms (GptOss case)
+            has_bias = fused_key + "_bias" in sd
+            if has_bias:
+                fused_bias_key = fused_key + "_bias"
+                gate_bias_key = gate_key + "_bias"
+                up_bias_key = up_key + "_bias"
+
+            # ---- split weights based on model type ----------------------
+            fused = sd[fused_key]  # [E, H, 2I]
             E, H, two_I = fused.shape
-            ffn_dim = two_I // 2
-            gate, up = fused.split(ffn_dim, dim=-1)  # views – no copy
 
-            experts = model_tmp.model.layers[layer_idx].feed_forward.experts
+            if is_gpt_oss:
+                # For GptOss, gate/up are interleaved: [gate0, up0, gate1, up1, ...]
+                gate = fused[..., ::2]  # [E, H, I] - even indices
+                up = fused[..., 1::2]  # [E, H, I] - odd indices
+            else:
+                # For standard MoE, gate/up are concatenated: [gate, up]
+                ffn_dim = two_I // 2
+                gate, up = fused.split(ffn_dim, dim=-1)  # views – no copy
+
+            # Copy weights to model
             experts.gate_proj.data.copy_(gate)
             experts.up_proj.data.copy_(up)
 
+            # Handle bias if present
+            if has_bias:
+                fused_bias = sd[fused_bias_key]  # [E, 2I]
+
+                if is_gpt_oss:
+                    gate_bias = fused_bias[..., ::2]  # [E, I] - even indices
+                    up_bias = fused_bias[..., 1::2]  # [E, I] - odd indices
+                else:
+                    ffn_dim = fused_bias.shape[-1] // 2
+                    gate_bias, up_bias = fused_bias.split(ffn_dim, dim=-1)
+
+                experts.gate_proj_bias.data.copy_(gate_bias)
+                experts.up_proj_bias.data.copy_(up_bias)
+
             # ---- update the state-dict so load_state_dict sees the right keys
             sd[gate_key] = gate
             sd[up_key] = up
 
+            if has_bias:
+                sd[gate_bias_key] = gate_bias
+                sd[up_bias_key] = up_bias
+
+            # Delete fused keys
             if delete_fused_key:
                 del sd[fused_key]
+                if has_bias:
+                    del sd[fused_bias_key]
 
-            logger.info(f"[layer {layer_idx:02d}] loaded gate_proj & up_proj from fused tensor  (shape {fused.shape})")
+            logger.info(f"[layer {layer_idx:02d}] loaded gate_proj & up_proj from fused tensor (shape {fused.shape})")
             transformed = True
 
         if hasattr(model, "language_model"):
             model.language_model = model_tmp
         else:
             model = model_tmp
+
         return model, transformed
 
 
-VLM_SPLIT_GATE_UP_WEIGHTS = {"QEffLlama4ForConditionalGeneration", "QEffLlama4ForCausalLM"}
+# Keep the existing list of supported models
+VLM_SPLIT_GATE_UP_WEIGHTS = {"QEffLlama4ForConditionalGeneration", "QEffLlama4ForCausalLM", "QEffGptOssForCausalLM"}

QEfficient/transformers/cache_utils.py

@@ -549,3 +549,122 @@ def update(
             ctx_v_out = torch.where(invalid_mask.unsqueeze(-1), torch.tensor(0.0, dtype=torch.float32), v_out)
             v_out = torch.where((is_sliding_layer & (position_ids.max() >= (layer_ctx_len - 1))), v_out, ctx_v_out)
         return k_out, v_out
+
+
+# This is a hack for now, until we get to merging this code with HybridCache class,
+# We don't really need to inherit transformers classes as their cache classes are made to work with pytorch and
+# ours are made to work with AIC
+class QEffHybridCacheForGPTOSS:
+    def __init__(self, config, batch_size, max_cache_len, sliding_window_len):
+        self.max_cache_len = max_cache_len
+        self.batch_size = batch_size
+        self.sliding_window_len = sliding_window_len
+        self.key_cache: List[torch.Tensor] = []
+        self.value_cache: List[torch.Tensor] = []
+
+    @classmethod
+    def from_legacy_cache(
+        cls, config, past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    ) -> "HybridCache":
+        """Converts a cache in the legacy cache format into an equivalent `DynamicCache`. Used for
+        backward compatibility."""
+        cache = cls(
+            config,
+            batch_size=past_key_values[0][0].shape[0],
+            max_cache_len=past_key_values[1][0].shape[2],
+            sliding_window_len=past_key_values[0][0].shape[2],
+        )
+        if past_key_values is not None:
+            for layer_idx in range(len(past_key_values)):
+                key_states, value_states = past_key_values[layer_idx]
+                cache.update(key_states, value_states, layer_idx)
+        return cache
+
+    def __len__(self):
+        """
+        Support for backwards-compatible `past_key_value` length, e.g. `len(past_key_value)`. This value corresponds
+        to the number of layers in the model.
+        """
+        return len(self.key_cache)
+
+    def get_seq_length(self, layer_idx: Optional[int] = 0) -> int:
+        """Returns the sequence length of the cached states. A layer index can be optionally passed."""
+        # TODO: deprecate this function in favor of `cache_position`
+        is_empty_layer = (
+            len(self.key_cache) == 0  # no cache in any layer
+            or len(self.key_cache) <= layer_idx  # skipped `layer_idx` and hasn't run a layer with cache after it
+            or len(self.key_cache[layer_idx]) == 0  # the layer has no cache
+        )
+        layer_seq_length = self.key_cache[layer_idx].shape[-2] if not is_empty_layer else 0
+        return layer_seq_length
+
+    def to_legacy_cache(self) -> Tuple[Tuple[torch.Tensor], Tuple[torch.Tensor]]:
+        """Converts the `DynamicCache` instance into the its equivalent in the legacy cache format. Used for
+        backward compatibility."""
+        legacy_cache = ()
+        for layer_idx in range(len(self)):
+            legacy_cache += ((self.key_cache[layer_idx], self.value_cache[layer_idx]),)
+        return legacy_cache
+
+    def update(
+        self,
+        key_states: torch.Tensor,
+        value_states: torch.Tensor,
+        layer_idx: int,
+        cache_kwargs: Optional[Dict[str, Any]] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        if len(self.key_cache) <= layer_idx:
+            self.key_cache.append(key_states)
+            self.value_cache.append(value_states)
+            k_out, v_out = key_states, value_states
+        else:
+            position_ids = cache_kwargs.get("position_ids")
+            is_sliding_layer = cache_kwargs.get("is_sliding")
+            sliding_window = cache_kwargs.get("sliding_window")
+            batch_index = cache_kwargs.get("batch_index", None)  # Check and fetch batch index value from the kwargs
+
+            if is_sliding_layer:
+                kv_position_ids = torch.where(position_ids == -1, position_ids, position_ids % sliding_window)
+            else:
+                kv_position_ids = position_ids
+
+            if batch_index is not None:
+                if torch.onnx.is_in_onnx_export():
+                    invalid_scatter_index = torch.iinfo(torch.int32).max
+                    scatter_position_ids = torch.where(kv_position_ids < 0, invalid_scatter_index, kv_position_ids)
+                else:
+                    scatter_position_ids = kv_position_ids
+                self.key_cache[layer_idx] = CtxScatterFuncCB.apply(
+                    self.key_cache[layer_idx], batch_index, scatter_position_ids, key_states
+                )
+                self.value_cache[layer_idx] = CtxScatterFuncCB.apply(
+                    self.value_cache[layer_idx], batch_index, scatter_position_ids, value_states
+                )
+            else:
+                self.key_cache[layer_idx] = CtxScatterFunc.apply(self.key_cache[layer_idx], kv_position_ids, key_states)
+                self.value_cache[layer_idx] = CtxScatterFunc.apply(
+                    self.value_cache[layer_idx], kv_position_ids, value_states
+                )
+
+            k_out, v_out = self.key_cache[layer_idx], self.value_cache[layer_idx]
+
+            # Original Gather
+            ctx_len = self.key_cache[layer_idx].shape[2]
+            ctx_indices = torch.arange(ctx_len)[None, None, ...]
+            gather_limit = position_ids.max(1, keepdim=True).values.unsqueeze(1)
+            invalid_mask = ctx_indices > gather_limit
+            if torch.onnx.is_in_onnx_export():
+                invalid_idx_value = torch.iinfo(torch.int32).max
+            else:
+                invalid_idx_value = 0
+            ctx_indices = torch.where(invalid_mask, invalid_idx_value, ctx_indices)
+
+            if batch_index is not None:
+                k_out = CtxGatherFuncCB.apply(k_out, batch_index, ctx_indices)
+                v_out = CtxGatherFuncCB.apply(v_out, batch_index, ctx_indices)
+            else:
+                k_out = CtxGatherFunc.apply(k_out, ctx_indices)
+                v_out = CtxGatherFunc.apply(v_out, ctx_indices)
+
+            v_out = torch.where(invalid_mask.unsqueeze(-1), torch.tensor(0.0, dtype=torch.float32), v_out)
+        return k_out, v_out

QEfficient/transformers/modeling_utils.py

@@ -185,6 +185,7 @@
     ]
 )
 
+# This is for supporting different seq_len for different layers for Sliding window attn, chunked attn etc.
 DYNAMIC_SEQ_LEN_SUPPORTED_MODEL_ARCH = {"gemma3", "llama4", "gemma3_text", "llama4_text"}
 
 # Define a transformers layers to QEff layers dictionary

QEfficient/transformers/models/gpt_oss/__init__.py

@@ -0,0 +1,6 @@
+# -----------------------------------------------------------------------------
+#
+# Copyright (c) Qualcomm Technologies, Inc. and/or its subsidiaries.
+# SPDX-License-Identifier: BSD-3-Clause
+#
+# ----------------------------------------------------------------------------