[ET-VK][patterns] Fuse torchao 4-bit quantized embedding to embedding_q4gsw

Pull Request resolved: #20381

TISO and other torchao-quantized models emit a `torchao.dequantize_affine -> aten.embedding` subgraph for their weight-only int4 quantized embedding. The existing `QuantizedEmbeddingMatch` only matches the `quantized_decomposed.embedding_4bit.dtype` fused op, so the torchao embedding never fused: its `dequantize_affine` const-folded to an fp32 weight, the resulting `aten.embedding` exceeded the buffer-element limit and fell back to CPU, and the fp32 constant bloated the serialized model.

This adds a separate `TorchAOQuantizedEmbeddingMatch` matcher that recognizes the torchao int4 `dequantize_affine -> aten.embedding` shape (qmin=-8/qmax=7, per-row group block_size `[1, G]`) and rewrites it to the existing `et_vk.embedding_q4gsw.default` op, repacking the unpacked int8 weight into the packed 4-bit layout. It asserts symmetric quantization (zero_point == 0, which the shader assumes) and guards against repacking a shared/tied weight more than once by recording the repack against the weight's state-dict FQN via `register_param_mutation` (a per-ExportedProgram registry on `ep._et_vk_param_modification_tags`); a second match on the same tied weight sees the recorded tag and skips the in-place repack. It is kept as a separate class from `QuantizedEmbeddingMatch` because the two dialects produce different graph shapes (one fused op vs a split dequant+gather), so a single class would only co-locate two disjoint parse paths.

On the en_US TISO backbone the embedding now delegates to Vulkan instead of falling back to CPU, and the serialized `.pte` drops from 418 MiB to 348 MiB.

This change was authored with Claude.
ghstack-source-id: 397529341
@exported-using-ghexport

Differential Revision: [D108457797](https://our.internmc.facebook.com/intern/diff/D108457797/)

Author: ssjia

backends/vulkan/patterns/BUCK

@@ -18,6 +18,7 @@ fbcode_target(_kind = runtime.python_library,
         "quantized_pixel_shuffle.py",
         "quantized_unary.py",
         "rms_norm.py",
+        "weight_packing_utils.py",
         "sdpa.py",
         "select_as_symint.py",
     ],

backends/vulkan/patterns/quantized_convolution.py

@@ -183,7 +183,7 @@ def find_quantized_convolution_patterns(
 
 
 @register_pattern_replacement("quantized_convolution")
-def make_q8ta_conv2d_custom_op(
+def make_q8ta_conv2d_custom_op(  # noqa: C901
     ep: ExportedProgram,
     graph_module: torch.fx.GraphModule,
     match: QuantizedConvolutionMatch,
@@ -249,9 +249,10 @@ def make_q8ta_conv2d_custom_op(
     # Need to make sure that OC dim is a multiple of 4 so that data load/stores are well
     # aligned with texel boundaries. Add padding to align to the next multiple of 4 if
     # needed.
-    utils.align_width_and_update_state_dict(
-        ep, match.weight_node, weight_tensor, force_update=True
-    )
+    if utils.register_param_mutation(ep, match.weight_node, "8 bit conv2d weight"):
+        utils.align_width_and_update_state_dict(
+            ep, match.weight_node, weight_tensor, force_update=True
+        )
     utils.align_width_and_update_state_dict(
         ep, match.weight_scales_node, weight_scales_tensor
     )

backends/vulkan/patterns/quantized_embedding.py

@@ -14,10 +14,18 @@
     register_pattern_detector,
     register_pattern_replacement,
 )
+from executorch.backends.vulkan.patterns.weight_packing_utils import (
+    pack_4bit_weight_tensor,
+)
 from executorch.exir import ExportedProgram
 from executorch.exir.dialects._ops import ops as exir_ops
 
 
+embedding_4bit_target = exir_ops.edge.quantized_decomposed.embedding_4bit.dtype
+embedding_target = exir_ops.edge.aten.embedding.default
+torchao_dequantize_affine_target = exir_ops.edge.torchao.dequantize_affine.default
+
+
 class QuantizedEmbeddingMatch(PatternMatch):
     def __init__(self, node: torch.fx.Node) -> None:
         self.anchor_node = node
@@ -65,51 +73,22 @@ def __init__(self, node: torch.fx.Node) -> None:
             self.scales_node = scales_node
             self.all_nodes.extend(arg_chain)
 
-        self.match_found = True
-
-
-embedding_4bit_target = exir_ops.edge.quantized_decomposed.embedding_4bit.dtype
-
-
-def _detect_tied_linear_weight(
-    ep: ExportedProgram,
-    weight_node: torch.fx.Node,
-    weight_tensor: torch.Tensor,
-) -> bool:
-    """Check if this embedding weight is tied to a linear weight.
-
-    The embedding weight is packed uint8 [vocab_size, embed_dim/2]. The linear
-    output weight may be stored as unpacked int8 [vocab_size, embed_dim]. If we
-    find a placeholder whose int8 values match our unpacked embedding values,
-    the weights are tied and we should use the linear packing to enable dedup.
-    """
-    vocab_size = weight_tensor.shape[0]
-    embed_dim = weight_tensor.shape[1] * 2
-
-    # Unpack embedding weight using embedding convention (high nibble first)
-    emb_high = (weight_tensor >> 4).to(torch.int8) - 8
-    emb_low = (weight_tensor & 0xF).to(torch.int8) - 8
-    emb_unpacked = torch.stack([emb_high, emb_low], dim=-1).reshape(
-        vocab_size, embed_dim
-    )
-
-    for node in ep.graph_module.graph.nodes:
-        if node.op != "placeholder" or node == weight_node:
-            continue
-
-        try:
-            candidate = get_param_tensor(ep, node)
-        except RuntimeError:
-            continue
-        if candidate is None:
-            continue
-        if candidate.shape != (vocab_size, embed_dim) or candidate.dtype != torch.int8:
-            continue
-
-        if torch.equal(emb_unpacked, candidate):
-            return True
+        # The weight placeholder stores values PACKED as uint8 [vocab,
+        # embed_dim / 2], so embed_dim is twice the inner dim. The op
+        # implementation requires that embed dim % 32 == 0 due to load/store
+        # granularity for the weight tensor; enforce that check now.
+        weight_val = (
+            self.weight_node.meta.get("val", None)
+            if isinstance(self.weight_node, torch.fx.Node)
+            else None
+        )
+        if not isinstance(weight_val, torch.Tensor) or weight_val.ndim != 2:
+            return
+        embed_dim = int(weight_val.shape[-1]) * 2  # packed, 2 values per byte
+        if embed_dim % 32 != 0:
+            return
 
-    return False
+        self.match_found = True
 
 
 @register_pattern_detector("quantized_embedding")
@@ -137,23 +116,25 @@ def replace_quantized_embedding_patterns(
     scales_tensor = get_param_tensor(ep, match.scales_node)
     assert scales_tensor is not None
 
-    is_linear_weight = _detect_tied_linear_weight(ep, match.weight_node, weight_tensor)
-
-    if is_linear_weight:
+    # The quantized_decomposed.embedding_4bit op (which is being replaced)
+    # already stores weights as packed uint8 [vocab, embed_dim / 2] (low nibble = odd,
+    # high nibble = even). However, in the Vulkan runtime 4-bit linear layers
+    # expect the reverse nibble packing (low nibble = even, high nibble = odd).
+    # In LLMs, where quantized embeddings are most frequently used, the embedding
+    # layer will share weights with the final LM head linear layer. For simplicity,
+    # always repack the weight tensor in the format expected by 4 bit linear layers;
+    # the runtime shader supports both the original and repacked packing formats
+    # for weights.
+    if utils.register_param_mutation(ep, match.weight_node, "4 bit linear weight"):
         # Repack using linear convention (low nibble = even, high nibble = odd)
         vocab_size = weight_tensor.shape[0]
         high = (weight_tensor >> 4).to(torch.int8) - 8
         low = (weight_tensor & 0xF).to(torch.int8) - 8
         unpacked = torch.stack([high, low], dim=-1).reshape(vocab_size, -1)
-        repacked = unpacked.to(torch.uint8) + 8
-        weight_tensor = repacked[:, 1::2] << 4 | repacked[:, ::2]
-        # Update the state dict with repacked tensor
-        original_weight = get_param_tensor(ep, match.weight_node)
-        if original_weight is not None:
-            for key, value in ep.state_dict.items():
-                if value.data_ptr() == original_weight.data_ptr():
-                    ep.state_dict[key] = weight_tensor
-                    break
+        weight_tensor = pack_4bit_weight_tensor(unpacked)
+        utils.align_width_and_update_state_dict(
+            ep, match.weight_node, weight_tensor, align_to=1, force_update=True
+        )
 
     # Compute group_size from weight and scales shapes
     embed_dim = weight_tensor.shape[1] * 2  # packed, 2 values per byte
@@ -169,7 +150,191 @@ def replace_quantized_embedding_patterns(
                 match.scales_node,
                 group_size,
                 match.indices_node,
-                is_linear_weight,
+                True,  # is_linear_weight
+            ),
+        )
+
+    embedding_q4gsw_node.meta["val"] = match.anchor_node.meta["val"]
+    match.anchor_node.replace_all_uses_with(embedding_q4gsw_node)
+
+
+class TorchAOQuantizedEmbeddingMatch(PatternMatch):
+    """Matches a torchao 4-bit weight-only quantized embedding and rewrites it
+    as a single et_vk.embedding_q4gsw.default node.
+
+    The recognized graph shape is a split torchao.dequantize_affine ->
+    aten.embedding, whose weight is unpacked int8 [vocab, embed_dim] with values
+    in [-8, 7]. This requires symmetric 4-bit signed quantization (quant_min=-8,
+    quant_max=7, zero_point=0) and per-row groupwise blocks (block_size=[1, G]),
+    which the runtime shader assumes via a fixed subtract-8 offset.
+    """
+
+    def __init__(self, node: torch.fx.Node) -> None:  # noqa: C901
+        self.anchor_node = node
+        self.match_found = False
+        self.all_nodes = [node]
+
+        # aten.embedding.default args: (weight, indices, *)
+        dequant_node = node.args[0]
+        self.indices_node = node.args[1]
+
+        if not isinstance(dequant_node, torch.fx.Node):
+            return
+        if dequant_node.target != torchao_dequantize_affine_target:
+            return
+
+        self.all_nodes.append(dequant_node)
+
+        # torchao.dequantize_affine args:
+        # (input, block_size, scale, zero_point, input_dtype, quant_min,
+        #  quant_max, ...)
+        block_size = dequant_node.args[1]
+        input_dtype = dequant_node.args[4] if len(dequant_node.args) > 4 else None
+        quant_min = dequant_node.args[5] if len(dequant_node.args) > 5 else None
+        quant_max = dequant_node.args[6] if len(dequant_node.args) > 6 else None
+
+        # The shader hardcodes the 4-bit signed offset (subtract 8), which
+        # corresponds to quant_min=-8, quant_max=7, zero_point=0.
+        if quant_min != -8 or quant_max != 7:
+            return
+
+        # Key off the dequant node's declared input_dtype, not the weight
+        # placeholder's live meta: a sibling match sharing the same (tied) weight
+        # may have repacked that placeholder in place (flipping it to packed
+        # uint8 [vocab, embed_dim / 2]), which would spuriously reject us here.
+        if input_dtype != torch.int8:
+            return
+
+        # block_size must be per-row groupwise: [1, group_size]
+        if not isinstance(block_size, (list, tuple)) or len(block_size) != 2:
+            return
+        if block_size[0] != 1:
+            return
+        self.group_size = int(block_size[1])
+
+        # Trace weight (args[0]) and scales (args[2]) to their placeholders. A
+        # placeholder-backed zero_point's symmetric (zero_point == 0)
+        # requirement is verified on the real tensor in the replacement
+        # function, where the ExportedProgram is available; checking the fake
+        # meta tensor here would trigger a data-dependent guard error.
+        weight_node, arg_chain = utils.trace_args_until_placeholder(
+            dequant_node.args[0]
+        )
+        if weight_node is None:
+            return
+        self.weight_node = weight_node
+        self.all_nodes.extend(arg_chain)
+
+        # Read embed_dim from the dequant node's float output meta, not the
+        # weight placeholder's meta: a tied weight may have been repacked in
+        # place by a sibling match (halving its inner dim), but this output meta
+        # is stable. Runtime shader requires embed_dim % 32 == 0 and the groups
+        # to tile the row exactly; reject otherwise rather than emit an op the
+        # runtime would abort on.
+        dequant_val = dequant_node.meta.get("val", None)
+        if not isinstance(dequant_val, torch.Tensor) or dequant_val.ndim != 2:
+            return
+        embed_dim = int(dequant_val.shape[-1])
+        if self.group_size <= 0 or embed_dim % self.group_size != 0:
+            return
+        if embed_dim % 32 != 0:
+            return
+
+        scales_node, arg_chain = utils.trace_args_until_placeholder(
+            dequant_node.args[2]
+        )
+        if scales_node is None:
+            return
+        self.scales_node = scales_node
+        self.all_nodes.extend(arg_chain)
+
+        # zero_point (args[3]) must be provably zero, since the shader hardcodes
+        # a subtract-8 offset that assumes symmetric quantization. Reject the
+        # match otherwise so the op falls back cleanly rather than miscomputing.
+        zero_point = dequant_node.args[3]
+        self.zero_point_node = None
+        if zero_point is None:
+            # Symmetric quant; zero_point == 0 is implied.
+            pass
+        elif isinstance(zero_point, torch.fx.Node):
+            zero_point_node, arg_chain = utils.trace_args_until_placeholder(zero_point)
+            if zero_point_node is None:
+                # Untraceable to a placeholder; cannot verify it is zero.
+                return
+            self.zero_point_node = zero_point_node
+            self.all_nodes.extend(arg_chain)
+        else:
+            # Inline scalar / list / tuple; verify the literal value(s) are zero.
+            values = (
+                zero_point if isinstance(zero_point, (list, tuple)) else [zero_point]
+            )
+            if any(v != 0 for v in values):
+                return
+
+        self.match_found = True
+
+
+@register_pattern_detector("torchao_quantized_embedding")
+def find_torchao_quantized_embedding_patterns(
+    node: torch.fx.Node,
+) -> Optional[TorchAOQuantizedEmbeddingMatch]:
+    if node.target != embedding_target:
+        return None
+
+    matched_pattern = TorchAOQuantizedEmbeddingMatch(node)
+    if matched_pattern.match_found:
+        return matched_pattern
+    return None
+
+
+@register_pattern_replacement("torchao_quantized_embedding")
+def replace_torchao_quantized_embedding_patterns(
+    ep: ExportedProgram,
+    graph_module: torch.fx.GraphModule,
+    match: TorchAOQuantizedEmbeddingMatch,
+):
+    # Always repack with the packing expected by 4 bit linear layers for
+    # simplicity. See replace_quantized_embedding_patterns() for more details
+    if utils.register_param_mutation(ep, match.weight_node, "4 bit linear weight"):
+        weight_tensor = get_param_tensor(ep, match.weight_node)
+        assert weight_tensor is not None
+
+        # The shader applies a fixed signed-4-bit offset (subtract 8), which
+        # assumes symmetric quantization (zero_point == 0). The None / inline
+        # literal cases were already proven zero in the matcher; a placeholder
+        # was committed to during matching, so its backing tensor must be
+        # fetchable and verifiable here.
+        if match.zero_point_node is not None:
+            zero_point_tensor = get_param_tensor(ep, match.zero_point_node)
+            if zero_point_tensor is None:
+                raise RuntimeError(
+                    "embedding_q4gsw: zero_point traced to placeholder "
+                    f"{match.zero_point_node.name!r} but its backing tensor "
+                    "could not be fetched to verify symmetric quantization "
+                    "(zero_point == 0)."
+                )
+            assert torch.all(
+                zero_point_tensor == 0
+            ), "embedding_q4gsw requires symmetric quantization (zero_point == 0)"
+
+        packed_weight = pack_4bit_weight_tensor(weight_tensor)
+
+        utils.align_width_and_update_state_dict(
+            ep, match.weight_node, packed_weight, align_to=1, force_update=True
+        )
+
+    group_size = match.group_size
+
+    with graph_module.graph.inserting_before(match.anchor_node):
+        embedding_q4gsw_node = graph_module.graph.create_node(
+            "call_function",
+            exir_ops.edge.et_vk.embedding_q4gsw.default,
+            args=(
+                match.weight_node,
+                match.scales_node,
+                group_size,
+                match.indices_node,
+                True,  # is_linear_weight
             ),
         )